AlphaEarth: Geospatial Foundation Model

Updated 3 July 2026

AlphaEarth is a global geospatial foundation model that fuses multi-modal satellite and ancillary data into standardized, 64-dimensional embeddings at 10 m resolution.
It employs a hybrid Transformer/convolutional STP encoder to integrate spatial, temporal, and measurement contexts, ensuring robust performance in diverse mapping tasks.
AlphaEarth streamlines global mapping by reducing feature engineering and preprocessing burdens, outperforming traditional hand-crafted and learned alternatives.

AlphaEarth is a geospatial foundation model and global embedding field released by Google DeepMind, designed to provide a unified, information-dense representation of Earth's surface from multi-modal satellite and ancillary sources. The AlphaEarth Foundations model outputs annual 64-dimensional embeddings at 10 m spatial resolution worldwide, enabling a standardized, analysis-ready input for a wide array of downstream tasks in environmental science, remote sensing, disaster management, urban analytics, and beyond. Its embedding fields have been shown to outperform traditional hand-crafted and learned alternatives on diverse global mapping, inference, and monitoring challenges.

1. Model Design, Sources, and Embedding Field Construction

AlphaEarth Foundations employs a Space–Time Precision (STP) encoder, a hybrid Transformer/convolutional architecture, to fuse spatial, temporal, and measurement context from multi-source remote sensing data (Brown et al., 29 Jul 2025). Primary inputs include Sentinel-2 (optical), Sentinel-1 (C-band SAR), Landsat 8/9 (optical/thermal), and other sources (LiDAR, climate, contextual text). At each ground pixel (10 m), all observations within a user-defined temporal window ([t_s, t_e)) are ingested. The encoder interleaves “space” self-attention (patch-level, ViT-style), “time” axial attention over {frame,patch} sequences (with sinusoidal time encodings), and “precision” convolutions (3×3 filters), with Laplacian-pyramid resampling maintaining information at multiple spatial scales.

The final layer produces a mean direction vector $\mu \in S^{63}$ , defining a von Mises–Fisher (vMF) distribution from which each embedding $u$ is sampled. Implicit decoders, each specialized to an input source, reconstruct held-out data (e.g., spectral channels, climate values) from $u$ and relevant metadata. A teacher–student contrastive framework ensures robustness to partial or missing modalities by aligning representations with inputs dropped at random.

Across training, a batch-uniformity objective maintains diversity in the representation space, while a CLIP-style contrastive loss aligns image and text (e.g., Wikipedia-derived) embeddings on the sphere. The total model has 480–1000M parameters, and post-training, Earth Engine delivers 64-channel, int8-quantized annual mosaics for 2017–2024 as “GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL” (Brown et al., 29 Jul 2025, Houriez et al., 15 Aug 2025).

2. Embedding Characteristics and Functional Organization

Each AlphaEarth embedding is a 64-dimensional real vector per 10 m pixel, lying close to the unit hypersphere under $\ell_2$ normalization (Benavides-Martinez et al., 8 Mar 2026). These vectors capture local and contextual signals from fused multi-modal time series, encoding surface reflectance, radar backscatter, topography, climate, and text-derived context.

Interpretability studies reveal a hierarchical dimension organization:

Specialist dimensions: strongly associated with specific land cover classes (e.g., mangroves, urban).
Low- and mid-generalist: shared across a few classes, capturing ecotone or transitional features.
High-generalist: reflecting broad environmental gradients (e.g., moisture, temperature).
Substantial redundancy: for operational land-cover classification, 2–12 of 64 dimensions suffice to retain 98% of baseline accuracy, enabling up to 80–90% reductions in runtime and storage per class (Benavides-Martinez et al., 8 Mar 2026).

Empirical tests (e.g., mean decrease in impurity, progressive ablation) validate that particular dimensions dominate inference for certain classes, and that high-level environmental structure can be reconstructed from a small subset of axes (Rahman, 10 Feb 2026, Benavides-Martinez et al., 8 Mar 2026).

3. Methodological Foundations and Downstream Integration

The embeddings are extracted as follows (Brown et al., 29 Jul 2025):

For each pixel, all available time-stamped input frames from each sensor are stacked.
STP encoder aggregates these across space and time, applying masking for self-supervision.
The output is a 64-dimensional embedding per ground pixel, per year.

These embeddings are consumed by a broad range of downstream models:

Shallow or deep classifiers/regressors for thematic or environmental variable mapping
Segmentation, object detection, or density estimation architectures (e.g., U-Net, Siamese U-Net)
Spatio-temporal models with embeddings as fixed spatial covariates (e.g., for EMS call forecasting (Aalaila et al., 1 Jul 2026))
Retrieval-augmented reasoning pipelines (embedding-based RAG with LLMs (Rahman, 10 Feb 2026, Rahman et al., 20 Apr 2026))

Importantly, these embeddings require no additional preprocessing by end-users: atmospheric correction, cloud masking, sensor registration, and principal component reduction are integrated upstream during AlphaEarth pipeline execution (Seydi, 9 Sep 2025, Ma et al., 30 Dec 2025).

4. Benchmarking and Representative Applications

AlphaEarth has demonstrated state-of-the-art or highly competitive performance on diverse global tasks, including:

Thematic mapping: Outperforms both learned (e.g., ViT-ImageNet, SatCLIP) and designed (CCDC harmonics, MOSAIKS) features on land cover, crop, and land-use prediction (Brown et al., 29 Jul 2025, Ma et al., 30 Dec 2025). For LCMAP land cover, achieves 0.78 balanced accuracy (vs. 0.70 best baseline); for Africa crop mask, 0.90 (vs. 0.82) (Brown et al., 29 Jul 2025).
Burned area mapping: Enables a bi-temporal Siamese U-Net to achieve 0.95 accuracy, F1-score 0.74, and IoU 0.6 on cross-continental benchmarks, with strong transferability and efficient generalization (Seydi, 9 Sep 2025).
Disaster mapping and change detection: High sensitivity to both abrupt and gradual land-surface change, with explicit bitemporal differencing integrated into models (Seydi, 9 Sep 2025, Cheng et al., 12 Jan 2026).
Hydrological and subsurface inference: Enhances ungauged basin flow prediction (median OOS NSE gain of 0.06–0.07 over traditional CAMELS attributes (Qu et al., 4 Jan 2026)) and subsurface temperature inference (test set $R^2$ up to 0.92) (Nakata et al., 16 Apr 2026).
Urban analytics: Supports mapping of SDG-aligned indicators, boundary- and density-based slum detection, local climate zone upscaling, and land-use/socioeconomic classification, outperforming or matching field-intensive baselines (Gong et al., 3 Apr 2026, Ko et al., 18 Jun 2026, Hou et al., 11 May 2026, Liu et al., 10 Oct 2025).

The embedding efficiency is highlighted by the finding that compact 64-d AlphaEarth vectors outperform PCA-reduced versions of much larger embeddings (e.g., Prithvi-768, Clay-1024) on spatially resolved urban tasks (Gong et al., 3 Apr 2026).

5. Limitations, Transferability, and Interpretability

While AlphaEarth embedding fields are robust across global domains, key limitations have been reported:

Domain drift: Distribution shift (ecoregion, climate, urban morphology) leads to reduced transferability for fine-grained or sensitive tasks (e.g., crop yield, slum density), with negative $R^2$ in out-of-region transfer scenarios (Ma et al., 30 Dec 2025, Hou et al., 11 May 2026).
Temporal granularity: Annual composites limit application in rapid-response or intra-seasonal tasks (e.g., real-time disaster, phenology).
Semantic opacity: Individual embedding axes typically lack direct physical interpretability; post-hoc attribution (SHAP, ablations) is needed to relate axes to land surface concepts (Benavides-Martinez et al., 8 Mar 2026).
Resolution ceiling: At 10 m, the embeddings cannot resolve intra-pixel structure in the most heterogeneous environments (e.g., positive-pixel $R^2$ for slum density is consistently negative at this scale (Hou et al., 11 May 2026)).
Dependence on training/supervision coverage: The value of AlphaEarth for operational mapping is highest in settings where local or regional label data exist; performance degrades for entirely novel regimes or where supervision is absent.

Nonetheless, studies show that with modest adaptation (e.g., lightweight multimodal alignment with POI data for urban tasks (Liu et al., 10 Oct 2025)), embeddings can be enriched with human-centric or functional semantics.

6. Impact on Geospatial Science and Practice

AlphaEarth’s embedding field paradigm represents a significant shift in remote sensing, environmental modeling, and scientific geoinformatics. It abstracts petabyte-scale Earth observation archives into a standardized, analysis-ready, and computationally efficient dense representation, allowing even classical ML models (logistic regression, random forests) to generalize beyond original label regions with minimal feature engineering (Houriez et al., 15 Aug 2025).

Major impacts include:

Enabling low-shot, global mapping in data-sparse regimes
Unifying input pipelines across thematic, physical, and functional mapping applications
Streamlining large-scale inference by reducing feature engineering and preprocessing burdens
Facilitating operational monitoring (burned area, floods, hydrology, SDGs) with consistent information layers
Supporting new forms of environmental reasoning with LLMs by enabling direct retrieval and dimensionally interpretable prompting (Rahman, 10 Feb 2026, Rahman et al., 20 Apr 2026)

The model’s widespread uptake in academic and commercial geospatial science, disaster management, planetary health, and urban/climate research reflects both its technical strengths and foundational role as a shared geospatial prior.

7. Future Directions and Ongoing Research

Multiple directions for extension and refinement have been identified:

Temporal upscaling: Production of higher-frequency embeddings (e.g., monthly or seasonal) to support intra-annual change detection, event forecasting, and crop/phenological analysis (Seydi, 9 Sep 2025, Ma et al., 30 Dec 2025).
Incorporation of additional data modalities: Integration of subsurface, anthropogenic, and high-resolution sociotechnical data to address context gaps and boost transferability (Qu et al., 4 Jan 2026).
Explainable AI and dimension selection: Systematic mapping of embedding dimensions to specific geophysical or human-functional concepts to reduce inference complexity and improve trust (Benavides-Martinez et al., 8 Mar 2026, Rahman, 10 Feb 2026).
Adaptation to new tasks: Lightweight alignment (e.g., with POIs or labels) to inject semantic context without retraining the full encoder, as demonstrated in urban applications (Liu et al., 10 Oct 2025).
Improved spatial regularization and error calibration: Mitigation of spatially structured errors and enhancement of uncertainty quantification (e.g., via Monte Carlo dropout or graph smoothing layers) (Hou et al., 11 May 2026, Narimani et al., 20 May 2026).
Operational deployment: Extension to Southern Hemisphere, boreal, and underrepresented biomes, and validation against new forms of ground truth as global label coverage expands (Seydi, 9 Sep 2025).

Ongoing evaluations continue to benchmark AlphaEarth and comparable foundation models on new science targets, operational products, and emerging planetary-scale monitoring challenges.