AlphaEarth Foundations Dataset

Updated 10 July 2026

AlphaEarth Dataset is a global geospatial representation where every 10m pixel is described by a 64-dimensional annual vector derived from diverse Earth observation sources.
It is distributed via Google Earth Engine as an analysis-ready feature store for tasks like mapping, forecasting, segmentation, and transfer learning.
The dataset leverages a Space–Time–Precision encoder and multi-modal training to provide robust context, despite limitations in intra-annual temporal resolution.

AlphaEArth Dataset, more commonly formalized as the annual embedding-field release of AlphaEarth Foundations, is a global geospatial representation in which each terrestrial 10 m pixel is encoded as a compact 64-dimensional annual vector summarizing multi-source Earth observation and contextual information over a chosen valid period (Brown et al., 29 Jul 2025). In operational usage, the term refers both to the foundation model and to the analysis-ready embedding product distributed for downstream use—most prominently through Google Earth Engine as “Satellite embedding v1 (annual)”—and consumed as fixed features for mapping, forecasting, regression, segmentation, retrieval, and transfer-learning pipelines (Houriez et al., 15 Aug 2025).

1. Naming, scope, and released product

The literature uses both “AlphaEarth” and “AlphaEArth” spellings, but papers that discuss the public release identify the formal dataset and Earth Engine asset with the AlphaEarth spelling (Seydi, 9 Sep 2025). A recurring distinction is between AlphaEarth Foundations as the upstream geospatial foundation model and the public embedding dataset as its practical output. Several downstream studies explicitly note that most users interact with the precomputed embeddings rather than the underlying model weights (Qu et al., 4 Jan 2026).

Aspect	Reported specification
Formal release	AlphaEarth Foundations / “Satellite embedding v1 (annual)”
Public access path	Google Earth Engine, `GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL`
Spatial coverage	Global terrestrial coverage, including minor islands, approximately between $\pm 82^\circ$ latitude
Spatial resolution	10 m
Representation size	64 channels per pixel per year
Public cadence in release paper	Annual layers for 2017–2024

The release paper describes a near-global terrestrial product with annual layers from 2017 through 2024 (Brown et al., 29 Jul 2025). Some downstream studies report adjacent distribution variants, including a Source Cooperative archive with annual data for 2018–2024 and application papers referring to availability through 2025 (Hamoudzadeh et al., 19 Feb 2026, Ko et al., 18 Jun 2026). This suggests distribution-specific versioning or later catalog extensions, rather than a single immutable packaging.

At the dataset level, AlphaEarth is presented as an analysis-ready representation intended to reduce dependence on bespoke per-task preprocessing. Multiple application papers adopt the embeddings as off-the-shelf inputs and emphasize that the downstream workflow often consists primarily of spatial extraction, aggregation, and a lightweight predictor rather than end-to-end retraining of the foundation model (Narimani et al., 20 May 2026).

2. Representation, temporal semantics, and source modalities

The core object is an embedding field $\mu(x,t_s,t_e)$ defined over space and a user-specified valid period $[t_s,t_e)$ (Brown et al., 29 Jul 2025). Each embedding is a 64-channel vector constrained to the unit sphere $S^{63}$ , intended to summarize the temporal evolution of multiple observations and contextual signals over that valid period. Public annual layers instantiate this mechanism at yearly cadence.

AlphaEarth assimilates a broad set of modalities during training. The release paper lists the following sources: Sentinel-2 L1C, Landsat-8/9 C2 Tier 1 TOA, Sentinel-1 GRD, ALOS-2/PALSAR-2 ScanSAR, GEDI LiDAR RH metrics, Copernicus DEM GLO-30, ERA5-Land monthly aggregates, GRACE mascons, NLCD land cover, geocoded Wikipedia articles, and GBIF taxa (Brown et al., 29 Jul 2025). Downstream papers typically summarize these as multi-source optical, radar, LiDAR/topography, climate, hydrology, and text-associated context (Ma et al., 30 Dec 2025).

The public dataset is annual, but the model itself is described as continuous-time conditioned. This matters because some applications treat the released layers as annual composites, whereas the underlying model formulation supports valid-period summaries beyond the annual public product. The release paper notes that some evaluations used monthly valid periods for evapotranspiration, even though the public dataset is annual (Brown et al., 29 Jul 2025). By contrast, application papers repeatedly note that the publicly consumed embeddings are annual summaries rather than near-real-time streams, which constrains tasks requiring fine intra-annual timing (Narimani et al., 20 May 2026).

The release also specifies a quantized public representation. Embeddings are stored as signed int8 values with power-law companding and dequantized via

$s = \frac{y}{127.5}, \qquad x = \operatorname{sign}(s)\,|s|^2$

after retrieval (Brown et al., 29 Jul 2025). One height-mapping study reports that a Source Cooperative distribution stores annual global embeddings as signed 8-bit integers in the range $[-128,127]$ and recommends a normalization to $[0,1]$ for local processing, while describing the Earth Engine catalog as accessible in double precision (Hamoudzadeh et al., 19 Feb 2026). Taken together, these reports indicate that the logical embedding field is stable, but serving formats can differ across distribution channels.

3. Model architecture and training objective

AlphaEarth Foundations is built around a Space–Time–Precision (STP) encoder with 15 STP blocks, each combining three parallel operators and learned Laplacian-pyramid exchanges (Brown et al., 29 Jul 2025). The three paths are a space path with ViT-like spatial self-attention at $1/16$ input size, a time path with axial self-attention conditioned on sinusoidal time codes, and a precision path with $3\times3$ convolutions at $1/2$ input size. Final features are projected through a von Mises–Fisher bottleneck with fixed $\mu(x,t_s,t_e)$ 0, and per-source implicit decoders reconstruct sensor-specific observations from the embedding, metadata, and normalized time code.

Training uses a weighted combination of reconstruction, uniformity, teacher–student consistency, and text-contrastive terms. The release paper gives the objective as

$\mu(x,t_s,t_e)$ 1

where $\mu(x,t_s,t_e)$ 2 is source-specific reconstruction loss, the second term enforces batch-wise uniformity on $\mu(x,t_s,t_e)$ 3, the third term is the teacher–student consistency penalty, and the fourth aligns embeddings with text embeddings through a CLIP-style loss (Brown et al., 29 Jul 2025).

The training corpus comprises 8,412,511 sequences from 5,145,244 sites and 3,047,520,515 frames across sources, covering approximately 1.1% of Earth’s land surface (Brown et al., 29 Jul 2025). Two model variants were trained, approximately 1B and 480M parameters, with deployment favoring the 480M model for efficiency. Training used 512 TPU v4, 100k steps, and batch size 256 sequences.

A central design goal is generalization without re-training. The release paper attributes this to continuous-time conditioning, source-specific decoders, robustness to missing or perturbed inputs via teacher–student training, batch-uniformity regularization, and text alignment, allowing the released embeddings to support simple downstream heads such as kNN and linear probes across heterogeneous tasks (Brown et al., 29 Jul 2025).

4. Distribution, access, and practical extraction workflows

The canonical public release is the Google Earth Engine collection GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL (Brown et al., 29 Jul 2025). The release paper also provides Zenodo links for processed evaluation datasets and training-site coordinates, but not downloadable model weights for local re-inference (Brown et al., 29 Jul 2025). A height-mapping study explicitly states that the AlphaEarth Foundations model weights are not open-source, even though the embeddings are accessible through public distribution channels (Hamoudzadeh et al., 19 Feb 2026).

Downstream usage is typically extraction rather than retraining. Studies aggregate embeddings to the task unit by zonal means, polygon clipping, basin averaging, or buffer-based sampling. Examples include averaging embeddings over field polygons for agricultural mapping (Narimani et al., 20 May 2026), averaging across basin masks and across years 2017–2024 for hydrological descriptors (Qu et al., 4 Jan 2026), extracting 10 m pixels within 2–5 km DHS buffers for health modeling (Nazir et al., 29 Apr 2026), and consuming embeddings directly on a 25×25 quadrature grid for spatio-temporal point-process forecasting (Aalaila et al., 1 Jul 2026).

Preprocessing is notably heterogeneous and task-dependent. Some papers state that no additional normalization was required because AlphaEarth bands were already bounded or analysis-ready (Narimani et al., 20 May 2026); others apply per-dimension standardization before cosine similarity or predictive modeling (Qu et al., 4 Jan 2026). Spatial harmonization is common: one landslide study resampled AlphaEarth rasters to 30 m × 30 m to match conventional conditioning factors (Cheng et al., 12 Jan 2026), while a subsurface study intentionally sampled the product at an effective 1 km scale in Earth Engine to stabilize centroid-level features (Nakata et al., 16 Apr 2026).

This use pattern has made AlphaEarth functionally similar to a large, annual geospatial feature store. Application papers repeatedly describe the practical pipeline as: query annual embeddings for the target year, aggregate or align to the study geometry, and fit a comparatively lightweight downstream model (Ma et al., 30 Dec 2025).

5. Downstream applications and empirical behavior

AlphaEarth embeddings have been applied across a wide spectrum of tasks, from classical land-cover mapping to point-process forecasting, subsurface inference, hydrology, public health, and urban analytics. Reported outcomes are highly task-specific, but several studies provide precise quantitative evaluations.

Domain	Configuration	Reported outcome
Spatio-temporal EMS forecasting	LGCP with AlphaEarth as linear spatial context across eight held-out regions	Mean multiplicative density ratio 5.42× at one week and 2.37× at two weeks; gains remain positive at longer histories (Aalaila et al., 1 Jul 2026)
Tomato cropping systems	U-Net on annual 64-band chips from 9,484 California fields	99.19% pixel accuracy, 99.04% F1, 98.11% IoU, 99.02% chip accuracy (Narimani et al., 20 May 2026)
Prediction in ungauged basins	LSTM with AlphaEarth basin descriptors on CAMELS-US	OOS median NSE = 0.612 vs 0.553 for CAMELS attributes; donor-basin selection peaks at 0.75 for selected basins at $\mu(x,t_s,t_e)$ 4 (Qu et al., 4 Jan 2026)
Urban indicator prediction	64-d embeddings across six U.S. metropolitan areas, 2020–2023	Global $\mu(x,t_s,t_e)$ 5 exemplars: drive alone 0.74, transit 0.72, obesity 0.69, log violent crime 0.48 (Gong et al., 3 Apr 2026)
Landslide susceptibility	CNN1D/CNN2D/ViT against conventional conditioning factors	F1-score gains of approximately 4%–15% and AUC increases of about 0.04–0.11 (Cheng et al., 12 Jan 2026)
Agricultural benchmarking	Yield, tillage, and cover-crop tasks in the U.S.	Strong local performance, but weaker spatial transfer; soybean US $\mu(x,t_s,t_e)$ 6Argentina AEF overall $\mu(x,t_s,t_e)$ 7 vs RS baselines $\mu(x,t_s,t_e)$ 8– $\mu(x,t_s,t_e)$ 9 (Ma et al., 30 Dec 2025)

A clear empirical pattern is that AlphaEarth is particularly effective when downstream tasks benefit from persistent landscape context and when labels are sparse or costly. The emergency-medical-services forecasting study is explicit on this point: with only 1–2 weeks of history, adding AlphaEarth to a fixed log-Gaussian Cox process produced approximately 2–6× multiplicative predictive-density gains in held-out regions (Aalaila et al., 1 Jul 2026). Hydrology and public-health studies report analogous benefits when embeddings are used as static or annually indexed descriptors of basin or cluster context (Qu et al., 4 Jan 2026, Nazir et al., 29 Apr 2026).

At the same time, several benchmarks show that the dataset’s utility is not uniform. In agriculture, AlphaEarth-based models were competitive with purpose-built remote-sensing features for local training, but exhibited limited spatial transferability, especially across ecoregions and countries (Ma et al., 30 Dec 2025). In fine-scale Local Climate Zone mapping, AlphaEarth simplified preprocessing and achieved strong performance, yet was consistently below TESSERA and showed larger sensitivity to cross-year transfer (Ko et al., 18 Jun 2026).

6. Interpretability, geometry, and recurring limitations

A substantial secondary literature analyzes the internal structure of AlphaEarth embeddings rather than using them only as predictors. One interpretability study over 12.1 million CONUS samples reports that the full 64-dimensional space can reconstruct many environmental variables with high fidelity: 12 of 26 variables exceed $[t_s,t_e)$ 0, and temperature and elevation approach $[t_s,t_e)$ 1; the same study finds a mean spatial block-cross-validation gap of $[t_s,t_e)$ 2 and mean inter-year correlation $[t_s,t_e)$ 3 for dimension–variable profiles (Rahman, 10 Feb 2026). This establishes that AlphaEarth dimensions are not arbitrary numerical codes, even if many applications treat them as latent and uninterpreted.

A separate geometric analysis characterizes the embedding manifold as strongly non-Euclidean. It reports global effective dimensionality PR $[t_s,t_e)$ 4 from 64 raw dimensions, local intrinsic dimensionality of approximately 10, weak local–global alignment with mean $[t_s,t_e)$ 5, and substantial tangent-space rotation, with 84% of locations exceeding $[t_s,t_e)$ 6 (Rahman et al., 20 Apr 2026). The same study finds that retrieval works substantially better than compositional vector arithmetic, and that local geometry predicts retrieval coherence with $[t_s,t_e)$ 7. This implies that AlphaEarth is more naturally used as a retrieval space than as a globally linear concept space.

A third line of work argues that dimensions can be organized along a functional hierarchy. In global land-cover experiments, accurate classification at 98% of baseline performance could be achieved with as few as 2 to 12 of the 64 dimensions, depending on the class, and dimensions were grouped into specialist, low-generalist, mid-generalist, and high-generalist categories based on their class associations (Benavides-Martinez et al., 8 Mar 2026). This suggests substantial redundancy in the embedding space and motivates dimension selection for operational inference.

Recurring limitations are equally clear. Across applied papers, the most frequently noted caveats are annual temporal granularity, limited time sensitivity for in-season tasks, weak direct semantics of individual channels, and variable spatial transferability (Ma et al., 30 Dec 2025, Ko et al., 18 Jun 2026, Cheng et al., 12 Jan 2026). Some studies also report tile gaps or incomplete coverage in particular locations (Ko et al., 18 Jun 2026). The embeddings encode surface context rather than direct subsurface or causal variables, so successful inference in domains such as site characterization or geothermal mapping remains indirect and dependent on robust validation design (Nakata et al., 16 Apr 2026). A plausible implication is that AlphaEarth is most reliable when used as a high-capacity contextual descriptor that complements, rather than universally replaces, domain-specific covariates and task-aware evaluation protocols.