cs.LGMar 8, 2026

What on Earth is AlphaEarth? Hierarchical structure and functional interpretability for global land cover

Authors: Ivan Felipe Benavides-Martinez, Justin Guthrie, Jhon Edwin Arias, Yeison Alberto Garces-Gomez, Angela Ines Guzman-Alvis, Cristiam Victoriano Portilla-Cabrera, Somnath Mondal, Andrew J. Allyn, 

Organizations: Artificial Intelligence for Climate and Sustainability The Institute for Experiential Artificial Intelligence Northeastern University, Portland, ME, USA · Gulf of Maine Research Institute Portland, ME, USA · Sustainability and Data Sciences Laboratory Northeastern University, Boston, MA, USA · School of Engineering and Architecture Universidad Católica de Manizales Manizales, Colombia · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Geospatial foundation models generate high-dimensional embeddings that achieve strong predictive performance, yet their internal organization remains obscure, limiting their scientific use. Recent interpretability studies relate Google AlphaEarth Foundations (GAEF) embeddings to continuous environmental variables, but it is still unclear whether the embedding space exhibits a functional or hierarchical organization, in which some dimensions act as specialized representations while others encode shared or broader geospatial structure. In this work, we propose a functional interpretability framework that reverse-engineers the role of embedding dimensions by characterizing their contribution to land cover structure from observed classification behavior. The approach combines large-scale experimentation with a structural analysis of embedding-class relationships based on feature importance patterns and progressive ablation. Our results show that embedding dimensions exhibit consistent and non-uniform functional behavior, allowing them to be categorized along a hierarchical functional spectrum: specialist dimensions associated with specific land cover classes, low- and mid-generalist dimensions capturing shared characteristics between classes, and highgeneralist dimensions reflecting broader environmental gradients. Critically, we find that accurate land cover classification (98% of baseline performance) can be achieved using as few as 2 to 12 of the 64 available dimensions, depending on the class. This demonstrates substantial redundancy in the embedding space and offers a pathway toward significant reductions in computational cost. Together, these findings reveal that AlphaEarth embeddings are not only physically informative, but also functionally organized into a hierarchical structure, providing practical guidance for dimension selection in operational classification tasks.

Figures & tables

Figure 1: Multi-modal fusion at scale, left) data containing the features, and right) its representation in an embedding.
Figure 2: Flowchart of the methodological framework: from embeddings to functional semantics. The pipeline integrates (1) ESA WorldCover 2020 categorical labels and AEF 64-dimensional embeddings as inputs, (2) a massive experimental framework executing more than 130,000 binary classification experiments using Random Forest, Gradient Boosting, XGBoost, and LightGBM with progressive feature ablation across the top 1 to 30 embedding dimensions, (3) structural mapping via an association matrix and functional taxonomy, and (4) an interactive embedding universe visualization connecting the latent space to semantic space.
 impA01impA02impA03impA04⋯\cdotsimpA64
Bare/sparse vegetation0.02670.02230.03500.0167⋯\cdots0.0409
Built-up0.00440.10920.04780.0057⋯\cdots0.0040
Cropland0.01630.02660.05640.0369⋯\cdots0.0179
Grassland0.02110.03150.05520.0270⋯\cdots0.0161
Herbaceous wetland0.01770.04490.02690.1090⋯\cdots0.0199
Mangroves0.04460.01030.01120.0162⋯\cdots0.0093
Table 1: Excerpt of the association matrix quantifying the normalized frequency with which each embedding dimension ranked among the two most important features by MDI across all experiments targeting each land cover class. Values range from 0 to 1, where higher values indicate stronger discriminative relevance of that dimension for the corresponding class
Figure 3: Embedding Fingerprint plot illustrating the functional dimensional signature of each land cover class at the 98%98\% performance recovery threshold. Analogous to a genomic fingerprint, where the presence and absence of bands at specific loci encodes biological identity, each row represents the characteristic pattern of exclusive (blue) and shared (pink) embedding dimensions that together define the discriminative identity of a land cover class within the AEF embedding space. Exclusive dimensions appear in only one class’s minimum subset, while shared dimensions — encompassing low-, mid-, and high-generalists — contribute to the classification of two or more classes. Gray positions indicate dimensions not required to achieve the 98%98\% threshold for that class. Land cover classes are ordered from top to bottom by increasing spectral complexity, quantified as the mean pixel-wise variance across Sentinel-2 spectral bands computed over 100 randomly selected images per class after standard atmospheric and radiometric corrections. Spectrally homogeneous classes such as Snow/ice and Water — which exhibit low inter-pixel variability due to their uniform reflectance signatures — appear at the bottom, while spectrally heterogeneous classes such as Built-up — whose radiometric response varies substantially across rooftops, roads, and impervious surfaces — appear at the top, reflecting the increasing number of embedding dimensions required to characterize them.
Figure 4: Conceptual structure of the embedding space representation, illustrating the planet-moon organization of the embedding universe. Blue circles represent land cover classes (planets), while green circles represent exclusive embedding dimensions (moons) whose color intensity encodes association strength — dark green indicating strong discriminative relevance, medium green indicating moderate relevance, and light green indicating weak relevance for that class. Gold diamonds represent shared embedding dimensions that connect multiple land cover classes, positioned at the geometric centroid of their associated classes and reflecting ecotonal or spectrally ambiguous regions where land cover boundaries overlap. This conceptual layout serves as the interpretive basis for the embedding universe visualization implemented in the interactive dashboard ( Guthrie and Benavides, 2025b ) and shown in Figure 6 .
Figure 5: Progressive ablation curves and minimum embedding dimensions by land cover class. Panels (a) and (b) show eight classification metrics as a function of MDI-ranked embedding dimensions for Water and Shrubland respectively (red = mean across experiments; blue dashed = 64-dimension baseline; green dotted = tipping point at 98%98\% baseline recovery). Water requires as few as 2 dimensions while Shrubland requires up to 12, reflecting differences in spectral distinctiveness. Panel (c) summarizes the minimum dimensions required to achieve 98%98\% of baseline performance across all 11 ESA WorldCover 2020 land cover classes.
Exclusive Embedding DimensionAssociated Land Cover ClassDistinctive trait that can justify an exclusive embedding
A09, A35Built-upUrban surfaces and infrastructure: roofs, roads, concrete. Their radiometric signature is dominated by artificial materials with high SWIR reflectance compared to NIR and strong Sentinel-1 backscatter. No other class combines this angular geometry and artificial-material spectrum, which justifies specific embeddings.
A12, A50CroplandAnnual crops sown and harvested within a 12-month cycle, with very regular phenology (well-marked greenness peaks), geometric patterns (plots, furrows) and often intensive management. Although they can be confused with grasslands, certain regions of the feature space (e.g. very homogeneous or highly irrigated crops) carry signatures specific to cropland, explaining their exclusive embeddings.
A05, A27MangrovesEvergreen woody vegetation tolerant to salinity in the intertidal zone. They show dense canopy like the Tree cover class, but with coastal linear patterns, water mixed between crowns and specific NIR/SWIR signatures due to permanent waterlogging. This combination (dense forest, brackish water and coastal context) is unique to mangroves.
A04, A11, A25, A29Herbaceous wetlandAreas dominated by herbaceous vegetation that is permanently or regularly flooded (fresh, brackish or saline water). They combine moderate–high NDVI with very high water indices (NDWI, etc.) and floodplain textures; no other type combines herbaceous cover and long-lasting flooding in this way.
A18, A21, A26ShrublandZones dominated by natural shrubs ( ≥\geq 10% cover), typical of semi-arid biomes. Although they mix with grassland and bare/sparse vegetation, subtypes where shrub height, density and spatial pattern (e.g. dense Mediterranean shrublands) generate signatures distinct enough to occupy an exclusive subspace.
A16, A23Tree coverRegions dominated by trees ( ≥\geq 10% cover), including plantations, with very high NIR, strong shadows and closed canopy textures. This 3-D structure and phenological persistence (especially in evergreen forests) clearly separate them from shrubland and grassland, allowing embeddings to specialise in representing “dense forests”.
Table 2: Exclusive embedding dimensions and their associated ESA WorldCover 2020 land cover classes, listing dimensions that appear solely in one class’s minimum subset alongside a biophysical interpretation of the distinctive spectral, structural, or phenological properties that justify their exclusive association.
Embedding DimensionAssociated ESA Land CoversCategorizationIdentifierCommon trait (ecological–spectral synthesis)
A07Cropland, Moss/lichenLow-generalistDiscontinuous vegetationAreas with low and discontinuous vegetation over mineral soil: croplands in fallow or early phenological stages and lichen mats show a soil and sparse vegetation mixture with high visible albedo and marked seasonal variability.
A08Built-up, Moss/lichenLow-generalistSparse vegetation over bright surfacesBright, sparsely vegetated substrates such as roofs, concrete and rock with lichens share low chlorophyll signal and rough textures that the model groups as mineral surfaces with minimal vegetation.
A15Moss/lichen, Snow/iceLow-generalistHigh latitude and high albedoHigh-latitude or high-altitude cold environments where snow/ice and lichen mats over rock produce bright visible reflectance and spatially homogeneous cryogenic landscapes.
A28Grassland, Tree coverLow-generalistDense green vegetationLandscapes dominated by green vegetation with strong NIR response where the main difference is canopy density; the embedding captures forest–grassland mosaics with intermediate canopy fraction.
A34Shrubland, Tree coverLow-generalistWoody perennialWoody perennial vegetation with relatively high biomass; shrublands and forests differ in height but share canopy textures and similar spectral and SAR responses.
A36Bare/sparse, Herbaceous wetlandLow-generalistSoil-moisture gradientTransition from dry bare soil to wet herbaceous substrates where exposed mud and seasonal wetlands resemble moist bare surfaces, capturing gradients of soil moisture.
Table 3: Shared embedding dimensions and their functional interpretation, grouped by generalist category (low-, mid-, and high-generalist) according to the number of land cover classes in whose minimum subset they appear. For each dimension, the associated ESA WorldCover 2020 classes and a common ecological–spectral trait are reported, inferred through the two-stage interpretive process described in Section 2.6.
Figure 6: “Embedding Universe” visualization at a 98% baseline performance recovery threshold, generated using the interactive Dashboard ( Guthrie and Benavides, 2025b ) . Green nodes represent specialist dimensions (darker = stronger association); gold nodes represent shared dimensions. See Section 2.5 for a full description of the visual encoding. To interact with this plot, access this link .
Supplementary Figure 1: Embedding frequency distributions by land cover class. For each class, embedding dimensions are ranked by frequency of appearance in the top 2 most important positions across 10,000 experimental runs. Green bars indicate dimensions within the tipping point threshold; blue bars indicate remaining dimensions. Generated using the interactive Dashboard ( Guthrie and Benavides, 2025b ) .
Supplementary Figure 2: Grid-based geographic performance heatmap of classification accuracy across global experiments. Each cell represents the mean classification accuracy aggregated across all experiments conducted within that geographic grid cell. Red cells indicate higher accuracy (>90%), orange cells indicate moderate accuracy (80–90%), and blue cells indicate lower accuracy (<80%). Generated using the interactive Dashboard ( Guthrie and Benavides, 2025b ) .
More in this topic

Geospatial Foundation Models

Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.

See all 122 papers

Explore similar work

CardsList
Apr 20, 2026

Characterizing AlphaEarth Embedding Geometry for Agentic Environmental Reasoning

Earth observation foundation models encode land surface information into dense embedding vectors, yet the geometric structure of these representations and its implications for downstream reasoning remain underexplored. We characterize the manifold geometry of Google AlphaEarth's 64-dimensional embeddings across 12.1 million Continental United States samples (2017--2023) and develop an agentic system that leverages this geometric understanding for environmental reasoning. The manifold is non-Euclidean: effective dimensionality is 13.3 (participation ratio) from 64 raw dimensions, with local intrinsic dimensionality of approximately 10. Tangent spaces rotate substantially, with 84% of locations exceeding 60\textdegree{} and local-global alignment (mean∣cos⁡θ∣=0.17|\cosθ| = 0.17) approaching the random baseline of 0.125. Supervised linear probes indicate that concept directions rotate across the manifold, and compositional vector arithmetic using both PCA-derived and probe-derived directions yields poor precision. Retrieval instead produces physically coherent results, with local geometry predicting retrieval coherence (R2=0.32R^2 = 0.32). Building on this characterization, we introduce an agentic system with nine specialized tools that decomposes environmental queries into reasoning chains over a FAISS-indexed embedding database. A five-condition ablation (120 queries, three complexity tiers) shows that embedding retrieval dominates response quality (μ=3.79±0.90μ= 3.79 \pm 0.90 vs.\ 3.03±0.773.03 \pm 0.77 parametric-only; scale 1--5), with peak performance on multi-step comparisons (μ=4.28±0.43μ= 4.28 \pm 0.43). A cross-model benchmark show that geometric tools reduce Sonnet 4.5's score by 0.12 points but improve Opus 4.6's by 0.07, with Opus achieving higher geometric grounding (3.38 vs.\ 2.64), suggesting that the value of geometric characterization scales with the reasoning capability of the consuming model.

Mashrekur Rahman, Samuel J. Barrett, Christina Last

Dartmouth Libraries, Dartmouth College, Hanover, 03755, NH, USA. · LGND AI, Canarias, Spain. · TipplyAI, London, UK.

Geospatial Foundation ModelsGeospatial ReasoningManifold LearningLLM Agents
AskPDF
Aug 4, 2026

Earth Embeddings

Earth observation is moving from foundation models that users must run themselves toward embedding products that package model feature outputs as reusable data without needing to download and process the imagery used to generate them. Earth embeddings are vectors that summarize locations, image patches, or pixels, letting users analyze compact features instead of repeatedly training or running large models on raw satellite imagery. This chapter explains the main types of Earth embeddings, from implicit location encoders to explicit patch and pixel products, and compares their coverage, resolution, dimensionality, storage cost, licenses, and reproducibility. We review their use in land cover and crop mapping, ecological and hazard modeling, socioeconomic prediction, and semantic search, with evidence on when embeddings improve on conventional features and when pooling, fusion, or spatial transfer limit performance. Two case studies show practical workflows for similarity search and land cover mapping. We close with guidance for choosing, evaluating, storing, compressing, and publishing embeddings, and with open problems in oceanic and atmospheric coverage, uncertainty, and benchmarking.

Adam J. Stewart, Heng Fang, Isaac A. Corley +1

Chair of Data Science in Earth Observation, Technical University of Munich, Munich, Germany · Munich Center for Machine Learning, Munich, Germany · KTH Royal Institute of Technology, Stockholm, Sweden +1

Geospatial Representation LearningLand Cover ClassificationRemote SensingEarth Science
AskPDF
May 18, 2026

Better Together: Evaluating the Complementarity of Earth Embedding Models

Earth embedding models transform Earth observation data into embeddings uniquely tied to locations on the Earth's surface. These models are typically evaluated in isolation, comparing the downstream task performance across different Earth embeddings. However, spatially aligned embeddings can naturally be fused, providing richer information per location, a capability that isolated evaluations fail to capture. We therefore propose assessing Earth embeddings by their complementarity: the performance gain of fused embeddings over the best single-model baseline. To operationalise this, we introduce an embedding complementarity index applicable to any embedding and task, and evaluate four Earth embedding models (AlphaEarth, Tessera, GeoCLIP, SatCLIP) in isolation, in all pairs, and jointly across six downstream tasks. Fused embeddings outperform the best single model in four out of six tasks, confirming that single-embedding evaluations often underestimate Earth embedding capabilities. Complementarity proves both task- and location-dependent. Further, for a land cover regression task, we find that complementarity is partially determined by the spatial scale of land cover classes. Complementarity reframes Earth embeddings: the greatest future gains may come not from any single Earth embedding model, but from combinations that are better together.

Thijs L van der Plas, Jacob JW Bakermans, Vishal Nedungadi +3

Wageningen University, NL · University College London, UK · University of Bonn, Germany

Geospatial Foundation ModelsGeospatial Representation LearningRemote SensingFeature Fusion
AskPDF