AlphaEarth Geospatial Embeddings

• AlphaEarth Geospatial Embeddings are compact, 64-dimensional annual representations that integrate multi-sensor data into analysis-ready Earth features.
• They leverage a Space-Time-Precision encoder with self-attention, CNN processing, and Laplacian exchanges to fuse spatial, temporal, and multi-modal measurements.
• Empirical evaluations demonstrate improved performance in classification, regression, and change detection across diverse applications, though temporal granularity remains a limitation.

AlphaEarth Geospatial Embeddings are compact, annual, analysis-ready representations of the Earth’s surface produced by AlphaEarth Foundations, the geospatial embedding-field model introduced by Brown et al. They map multi-source Earth observation measurements and time context into a fixed 64-dimensional vector at approximately 10 m spatial resolution, and are intended to serve as universal downstream features rather than task-specific spectral indices or manually engineered descriptors. In the published literature, these embeddings are used with frozen encoders and lightweight downstream models for classification, regression, change detection, hydrology, agriculture, urban analytics, and environmental reasoning (Brown et al., 29 Jul 2025).

1. Representation and data model

Downstream papers refer to the same product as AlphaEarth Foundations embeddings, AlphaEarth Satellite Embeddings, or simply Earth embeddings. In the foundation formulation, the embedding field is a function

$z = f_\theta(x, y, t, M_{1:K}) \in \mathbb{R}^d,$

where $(x,y)$ denotes geospatial position, $t$ denotes temporal conditioning, and $M_{1:K}$ denotes multi-modal measurement streams; the released configuration uses $d=64$ and constrains the latent to the unit hypersphere $S^{63}$ (Brown et al., 29 Jul 2025). The model distinguishes a support period, which defines the interval of available input measurements, from a valid period $[t_s,t_e)$, which defines the summary interval requested by the user (Brown et al., 29 Jul 2025).

The minimal inference inputs in the foundation paper are sequences from Sentinel-2 L1C, Sentinel-1 GRD, and Landsat-8/9 L1C TOA. During training, reconstruction targets additionally include GEDI LiDAR, PALSAR-2, ERA5-Land, GRACE, GLO-30 DEM, NLCD, and geocoded text, so the learned embedding is explicitly multi-sensor, multi-source, and time-aware rather than a compressed optical image descriptor (Brown et al., 29 Jul 2025).

Operational papers emphasize that the 64 channels are not intended to be interpreted as individually meaningful spectral or physical bands. In the California tomato workflow, each 10 m pixel is represented by a 64-band embedding vector A00–A63, already normalized to $[-1,1]$, and interpreted jointly as a learned feature-space coordinate rather than as hand-interpretable variables (Narimani et al., 20 May 2026). Other access routes expose the same annual product in different storage forms: the foundation release uses 8-bit signed quantization with 64 bytes per pixel per year (Brown et al., 29 Jul 2025), while a height-mapping study reports double-precision Google Earth Engine assets and signed 8-bit Source Cooperative Cloud-Optimized GeoTIFFs in the range $[-128,127]$ (Hamoudzadeh et al., 19 Feb 2026).

Most downstream studies describe the public annual layers as covering 2017–2024 (Brown et al., 29 Jul 2025), although a Swiss Local Climate Zone benchmark describes global annual embeddings for 2017–2025 in its summary of Brown et al. (Ko et al., 18 Jun 2026). This temporal granularity is annual rather than near-real-time, a property that is central to both the strengths and the limitations of the representation.

2. Foundation architecture and training regime

AlphaEarth Foundations is built around a Space-Time-Precision encoder. The architecture uses repeated space, time, and precision operators connected by learned Laplacian-pyramid exchanges: a ViT-like spatial self-attention path at $L/16$, axial temporal self-attention conditioned on sinusoidal timecodes at $(x,y)$0, and 3×3 CNN processing at $(x,y)$1 (Brown et al., 29 Jul 2025). A single learned query summarizes the encoded sequence over the requested valid period, and the latent embedding is represented as the mean direction of a von Mises–Fisher distribution, which regularizes the compact 64-dimensional bottleneck (Brown et al., 29 Jul 2025).

Training combines four objectives. The first is reconstruction across multiple gridded sources; the second is batch uniformity, which encourages a uniform distribution on the hypersphere; the third is teacher–student consistency under structured source and timestamp dropout; and the fourth is CLIP-style text alignment against geocoded Wikipedia and GBIF descriptions. The reported loss weights are $(x,y)$2, $(x,y)$3, $(x,y)$4, and $(x,y)$5 (Brown et al., 29 Jul 2025).

The scale of pretraining is correspondingly large. The foundation paper reports 8,412,511 multi-source video sequences from 5,145,244 sites, 10,203,798 unique $(x,y)$6 rows, and 3,047,520,515 frames spanning 2017–2024, sampled over approximately 1.1% of global land area (Brown et al., 29 Jul 2025). Training used 56 hours on 512 TPU v4, a batch size of 256 sequences, 100k steps, and both approximately 1B-parameter and approximately 480M-parameter variants, with the smaller model selected for inference efficiency (Brown et al., 29 Jul 2025).

A defining architectural feature is the separation between sensor-specific nuisance variation and shared geophysical information. The implicit decoders reconstruct held-out targets at arbitrary timestamps inside the valid period, conditioned on source metadata and relative time. This means the embedding is trained not merely to compress one modality, but to remain predictive across modalities and across temporal interpolation or extrapolation settings (Brown et al., 29 Jul 2025).

3. Release format and operational use

The released product is an annual stack of 64-band global embedding layers served through Google Earth Engine under the Satellite Embedding V1 annual dataset (Brown et al., 29 Jul 2025). In practice, most applied studies use AlphaEarth exactly as published, without retraining the encoder, and build downstream workflows by either sampling per-pixel vectors or aggregating them over points, polygons, buffers, or regular grids.

In field-scale crop segmentation, the California tomato study clips annual 64-band chips to LandIQ polygons, aligns them to binary masks, and trains a 64-channel U-Net with masked binary cross-entropy plus soft Dice loss. Monte Carlo dropout is retained at inference and repeated $(x,y)$7 times per chip to estimate predictive mean and variance maps (Narimani et al., 20 May 2026). In hydrology, annual embeddings from 2017–2024 are averaged over basin footprints and years to obtain a single 64-dimensional basin descriptor, standardized per dimension, compared by cosine similarity, and concatenated with daily meteorological forcings before an MLP–LSTM rainfall–runoff model (Qu et al., 4 Jan 2026).

Other studies use equally lightweight aggregation schemes. The poverty-mapping study samples the precomputed raster within a 500 m radius around each DHS or GeoNames coordinate and averages the pixels to obtain a single 64-dimensional node feature for graph neural networks (Pettersson et al., 3 Nov 2025). The Malawi health-facility study aggregates pixel-level AlphaEarth features to 552 catchment polygons and feeds them to XGBoost models for routine health-indicator prediction (Metz et al., 29 Oct 2025). The Brazilian restoration study computes annual polygon means and compares them to mature secondary-forest references through cosine similarity trajectories (Heiman, 7 May 2026).

This operational pattern is one of the most consistent empirical properties of the AlphaEarth literature: the embeddings are treated as a fixed, analysis-ready interface between heterogeneous Earth observation data and small downstream heads such as ridge regression, random forests, XGBoost, U-Nets, LSTMs, MLPs, or GCNs, rather than as a model family that must be end-to-end fine-tuned for each task (Brown et al., 29 Jul 2025).

4. Empirical performance across application domains

In the foundation evaluation suite, frozen AlphaEarth embeddings plus lightweight heads consistently outperformed the baselines tested across 15 evaluations from 11 datasets. The reported average error-magnitude reduction relative to the next-best approach was approximately 23.9% in max-trial settings, approximately 10.4% in 10-shot settings, and approximately 4.18% in 1-shot settings (Brown et al., 29 Jul 2025). The same paper reports $(x,y)$8 for ASTER GED surface emissivity, $(x,y)$9 for OpenET monthly evapotranspiration, and balanced accuracy values up to 78.4% and 79.3% for directly supervised LCMAP land-cover and land-use change detection (Brown et al., 29 Jul 2025).

Agricultural results are strong but heterogeneous. In California, a U-Net trained on 4,742 tomato and 4,742 non-tomato fields achieved 99.19% pixel accuracy, 98.69% precision, 99.40% recall, 99.04% F1 score, 98.11% intersection over union, and 99.02% chip accuracy, with uncertainty consistently highest near field edges and lowest in field interiors (Narimani et al., 20 May 2026). In Togo, a Google Earth Engine Random Forest trained on AlphaEarth embeddings produced a 10 m cropland map with Overall Accuracy 0.859, User’s Accuracy 0.745, Producer’s Accuracy 0.745, and F1 Score 0.745 (Zvonkov et al., 4 Nov 2025). A broader U.S. agricultural benchmark found AEF competitive with purpose-built remote-sensing feature pipelines for local county-level yield prediction and some field-level tasks, but notably weaker in scale-transfer and in cross-region or cross-country transfer, including strongly negative soybean-yield $t$0 values in U.S.→Argentina transfer (Ma et al., 30 Dec 2025).

Hydrologic and hazard applications show similar task-dependent gains. Replacing hand-crafted CAMELS attributes with AlphaEarth basin embeddings increased median out-of-sample NSE from 0.553 to 0.612, while AEF-based donor selection was most effective for small-to-moderate donor sets and degraded when many dissimilar basins were added (Qu et al., 4 Jan 2026). In landslide susceptibility mapping across Taiwan, Hong Kong, and Emilia-Romagna, AlphaEarth-based models improved F1 by approximately 4% to 15% and AUC by 0.04 to 0.11 over conventional landslide conditioning factors (Cheng et al., 12 Jan 2026). In height inference from annual embeddings, U-Net++ achieved test $t$1, RMSE 16.42 m, and median difference $t$2 m, outperforming both standard U-Net and ridge regression (Hamoudzadeh et al., 19 Feb 2026). For shallow seismic site characterization, a hybrid XGBoost model using embeddings plus log-slope achieved RMSE 126.1 m/s and $t$3, versus RMSE 166.6 m/s and $t$4 for a covariate-only linear baseline (Nakata et al., 16 Apr 2026).

Urban and public-health studies broaden the scope further. Across six U.S. metropolitan areas and 14 neighborhood indicators, AlphaEarth achieved global test $t$5 values of 0.74 for %Drive alone, 0.72 for %Transit, 0.69 for %Obesity, 0.48 for log violent crime rate, and 0.44 for log median household income, while remaining more informative than 64-dimensional reductions of Prithvi and Clay (Gong et al., 3 Apr 2026). In slum monitoring across 12 cities and 69 city-year pairs, same-city cross-year training yielded median spatial F1 = 0.616 and $t$6, but positive-pixel $t$7 was consistently negative across all cities, indicating that most regression skill came from zero/non-zero discrimination rather than intra-pixel density modeling (Hou et al., 11 May 2026). In climate-sensitive health prediction, adding AlphaEarth increased Nigeria malaria test $t$8 from 0.201 to 0.245 and raised pooled childhood acute respiratory infection $t$9 from approximately 0.157 to approximately 0.206 across three tree-based estimators and 11 DHS countries (Nazir et al., 29 Apr 2026).

Not all comparative studies favor AlphaEarth over other embeddings. In fine-scale Swiss Local Climate Zone mapping, AlphaEarth was competitive but consistently outperformed by TESSERA, with test IoU ranges of 0.59–0.69 in multi-city transfer and 0.77–0.82 in a higher-resolution Bern setting, where TESSERA reached 0.82, AlphaEarth 0.81, and S1S2 0.77 (Ko et al., 18 Jun 2026). This comparative record is important because it shows that AlphaEarth is not uniformly dominant; it is a strong general-purpose representation whose relative advantage depends on the downstream problem, the label structure, and the competing feature family.

5. Interpretability, hierarchy, and embedding geometry

Interpretability studies have shifted the description of AlphaEarth from a purely black-box feature source to a physically and functionally structured latent space. Across 12.1 million CONUS samples and 26 environmental variables, 12 variables exceeded $M_{1:K}$0 when reconstructed from the full 64-dimensional embedding space, and temperature and elevation approached $M_{1:K}$1 (Rahman, 10 Feb 2026). The strongest reported dimension–variable pairings include A57 with annual precipitation, LAI, dew point, and elevation; A40 with daytime land-surface temperature; A48 with EVI, LAI, and NDVI; A26 with tree cover; and A00 with evapotranspiration, precipitation, and soil moisture (Rahman, 10 Feb 2026). These mappings remained spatially robust under block cross-validation with mean $M_{1:K}$2 and temporally stable across 2017–2023 with mean inter-year correlation $M_{1:K}$3 (Rahman, 10 Feb 2026).

A separate land-cover study argued that the 64 coordinates occupy a hierarchical functional spectrum rather than a flat, undifferentiated representation. It identified specialist dimensions associated with a single land-cover class, low- and mid-generalists associated with two or three classes, and high-generalists associated with broader environmental gradients (Benavides-Martinez et al., 8 Mar 2026). In that framework, accurate land-cover classification could achieve 98% of baseline performance with as few as 2 to 12 dimensions depending on class; 43 dimensions were interpreted through their classification behavior, 21 were not selected in any minimum subset, and pruning reduced classification time by 20%–80% (Benavides-Martinez et al., 8 Mar 2026).

Geometric analysis complicates any simple linear reading of this structure. A large-scale manifold study reported an effective dimensionality of 13.3 by participation ratio, mean local intrinsic dimensionality of approximately 10, mean local-global alignment $M_{1:K}$4 for PC1, and principal-angle rotations above 60° for 84% of adjacent tangent-space pairs (Rahman et al., 20 Apr 2026). Under these conditions, compositional vector arithmetic and globally reusable “concept directions” perform poorly, even when supervised linear probes obtain high predictive $M_{1:K}$5 (Rahman et al., 20 Apr 2026). Retrieval, by contrast, remains physically coherent, and local geometric features predict retrieval coherence with $M_{1:K}$6 (Rahman et al., 20 Apr 2026). This distinction between predictive linear probes and non-linear manifold geometry is one of the most consequential recent findings in the AlphaEarth literature.

6. Limitations, controversies, and extensions

Several limitations recur across the applied record. The public product used in most studies is annual rather than near-real-time, so intra-annual phenology and short-window management signals are only weakly represented (Narimani et al., 20 May 2026). This is explicitly visible in agriculture, where AEF embeddings were judged to have limited time sensitivity and weaker spatial transferability than remote-sensing feature pipelines for some yield-transfer settings (Ma et al., 30 Dec 2025), and in Swiss LCZ mapping, where transfer from one year to another remained an open challenge and Sentinel-1/2 composites were more stable than AlphaEarth under year-to-year deployment (Ko et al., 18 Jun 2026).

Generalization is also highly task-dependent. In hydrology, leave-one-cluster-out experiments produced aggregated NSE = 0.199 when clusters were defined by AEF embeddings, compared with 0.297 when clusters were defined by CAMELS attributes, which the authors attribute to a granularity–generalization trade-off: AEF produces more discriminative regimes, but those regimes can become harder to extrapolate across (Qu et al., 4 Jan 2026). In Brazilian Atlantic Forest restoration, embeddings separated land-use and land-cover classes, highlighted change vectors, and exposed outliers, but the signal could be noisy and did not improve restoration-strategy prediction relative to environmental and spectral features (Heiman, 7 May 2026). In slum density mapping, positive-pixel $M_{1:K}$7 remained negative across all cities despite good boundary discrimination, showing that 10 m annual embeddings do not automatically resolve intra-pixel socioeconomic gradients (Hou et al., 11 May 2026).

A recurrent criticism in urban work is that EO-driven embeddings encode physical morphology more strongly than human activity, spatial function, or socioeconomic semantics. AETHER addresses this limitation by aligning fixed AlphaEarth embeddings with POI text through symmetric InfoNCE losses. In Greater London, this increased land-use-classification F1 from 56.6 ± 0.5 to 60.7 ± 0.4 and reduced socioeconomic-mapping KL from 43.2 ± 0.6 to 33.0 ± 0.3, while keeping the AlphaEarth backbone frozen (Liu et al., 10 Oct 2025). This extension does not replace AlphaEarth; rather, it treats the original embedding field as a physically grounded prior that can be enriched by auxiliary modalities when the downstream task depends on function rather than appearance alone (Liu et al., 10 Oct 2025).

The published record therefore presents AlphaEarth Geospatial Embeddings as compact, globally consistent, annual latent descriptors that substantially reduce preprocessing burden and often support strong label-efficient performance. At the same time, they are not human-readable bands, not uniformly transferable across space or time, and not sufficient by themselves for every semantic or process-level task. Their most stable role in the literature is as an analysis-ready backbone to be paired with carefully chosen downstream models, validation protocols, and, where necessary, complementary semantic or domain-specific information.