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

W H A T O N E A RT H I S A L P H A E A RT H ? H I E R A R C H I C A L S T RU C T U R E A N D F U N C T I O N A L I N T E R P R E T A B I L I T Y F O R G L O B A L L A N D C O V E R A P R E P R I N T Iván Felipe Bena vides-Martínez ∗ Artificial Intelligence for Climate and Sustainability The Institute for Experiential Artificial Intelligence Northeastern Univ ersity , Portland, ME, USA Gulf of Maine Research Institute, Portland, ME, USA i.benavides@northeastern.edu Justin Guthrie † Sustainability and Data Sciences Laboratory Northeastern Univ ersity , Boston, MA, USA j.guthrie@northeastern.edu Jhon Edwin Arias ‡ School of Engineering and Architecture Univ ersidad Católica de Manizales Manizales, Colombia jhon.arias2@ucm.edu.co Y eison Alberto Garcés-Gómez § School of Engineering and Architecture Univ ersidad Católica de Manizales Manizales, Colombia ygarces@ucm.edu.co Angela Ines Guzman-Alvis ¶ Grupo de In vestigación en Recursos Hidrobiológicos Departamento de Ingeniería Univ ersidad Nacional de Colombia Palmira, Colombia aiguzmana@unal.edu.co Cristiam V ictoriano Portilla-Cabr era ∥ Grupo de In vestigación en Recursos Hidrobiológicos Departamento de Ingeniería Univ ersidad Nacional de Colombia Palmira, Colombia cvportillac@unal.edu.co Somnath Mondal ∗∗ Artificial Intelligence for Climate and Sustainability The Institute for Experiential Artificial Intelligence Northeastern Univ ersity , Portland, ME, USA s.mondal@northeastern.edu Andrew J . Allyn †† Gulf of Maine Research Institute Portland, ME, USA aallyn@gmri.org A uroop R. Ganguly ‡‡ Artificial Intelligence for Climate and Sustainability The Institute for Experiential Artificial Intelligence Northeastern Univ ersity , Portland, ME, USA a.ganguly@northeastern.edu A B S T R AC T ∗ https://orcid.org/0000-0002-1139-3909 † https://orcid.org/0009-0009-9133-6135 ‡ https://orcid.org/0009-0009-9968-8776 § https://orcid.org/0000-0002-9409-3652 ¶ https://orcid.org/0000-0002-5185-8950 ∥ https://orcid.org/0000-0003-4346-3972 ∗∗ https://orcid.org/0000-0001-8217-1361 †† https://orcid.org/0000-0002-1584-0198 ‡‡ https://orcid.org/0000-0002-4292-4856 W H AT ON E A RT H I S A L P H A E A RT H ? A P R E P R I N T Geospatial foundation models generate high-dimensional embeddings that achie ve strong predicti ve performance, yet their internal organization remains obscure, limiting their scientific use. Recent interpretability studies relate Google AlphaEarth Foundations (GAEF) embeddings to continuous en vironmental 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 inter- pretability frame work that re verse-engineers the role of embedding dimensions by characterizing their contribution to land co ver structure from observ ed classification behavior . The approach combines large-scale experimentation with a structural analysis of embedding–class relationships based on feature importance patterns and progressi ve ablation. Our results sho w 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 cov er classes, low- and mid-generalist dimensions capturing shared characteristics between classes, and high- generalist dimensions reflecting broader en vironmental gradients. Critically , we find that accurate land cov er classification ( 98% of baseline performance) can be achiev ed using as few as 2 to 12 of the 64 a vailable dimensions, depending on the class. This demonstrates substantial redundancy in the embedding space and offers a pathway toward significant reductions in computational cost. T ogether , these findings reveal that AlphaEarth embeddings are not only physically informati ve, but also functionally organized into a hierarchical structure, pro viding practical guidance for dimension selection in operational classification tasks. Keyw ords Geospatial F oundation Models · Earth Observ ation · Remote Sensing · V irtual Satellites · Explainable AI · Responsible AI · Feature Attrib ution · Interpretable Embeddings 1 Introduction Geospatial foundation models (GFMs) — part of a broader movement toward comprehensiv e world models that simulate the dynamics of Earth’ s interconnected systems — are transforming Earth observation research in fields such as ecology , infrastructure planning, and climate science (Bodnar et al., 2025; Editorial, 2025). These models function as "virtual satellites" capable of characterizing the Earth’ s surface and its dynamics with an unprecedented level of detail (Google, 2025). GFMs hav e emerged as a response to tw o classic bottlenecks in Earth observ ation: conv erting large v olumes of multi-source data into useful information and the limited generalization of task-specific models when they depend on high-quality labels (Editorial, 2025). In this context, Google AlphaEarth Foundations (GAEF) was proposed as a task-agnostic learned featurization approach that integrates multiple data sources, producing eight annual 64-dimensional embeddings at 10 m (2017 to 2024) to support operational mapping and learning with sparse labels (Brown et al., 2025; Editorial, 2025). GAEF addresses traditional limitations of remote sensing, such as source heterogeneity , differences in spatial and temporal resolution, and the difficulty of integrating multiple sensors by synthesizing large volumes of data from optical images, radar , LiD AR, climate v ariables, and other sources (Figure 1 left) into a unified and consistent representation of the planet (Houriez et al., 2025; Google, 2025). Howe ver , despite demonstrated ef fectiveness in various mapping and en vironmental modeling tasks (Brown et al., 2025; Ma et al., 2025), these models present a fundamental interpretability challenge: they encode en vironmental information into high-dimensional embeddings that compress complex geophysical patterns into abstract numerical vectors (Figure 1 right) Brown et al. (2025); Houriez et al. (2025). Sev eral criticisms hav e recently been raised about GAEF , arguing that, despite its strong benchmark performance, it functions as an opaque, annually updated 10 - m “black - box” representation whose mix ed - pixel embeddings, limited physical and human - centric interpretability , and dependence on biased upstream data constrain its reliability for high-stakes, fine-scale, and socio-ecological decision-making in real-world settings (Rahman, 2026; Liu et al., 2025). Unlike conv entional remote sensing products, where each band is associated with a measurable physical property such as spectral reflectance, temperature, or humidity (Seydi, 2025), these embeddings represent latent features that capture complex Earth surface patterns without an e xplicit biogeophysical interpretation (Liu et al., 2025). Their values, usually defined in normalized, high-dimensional spaces, do not offer a direct interpretation of the en vironmental processes they represent, creating a significant gap between the predicti ve capacity of the models and their scientific understanding (Rahman, 2026; Liu et al., 2025). This limitation raises the need to de velop methodological frame works that allo w these embeddings to be interpreted in terms of trackable and reproducible physical variables and en vironmental processes, facilitating their integration into scientific decision-making. W ithout such frameworks, the utility of embeddings is limited to predicti ve performance, with little capacity to explain what en vironmental information is encoded or ho w they relate to the biogeophysical characteristics of the Earth’ s surface. Bridging this gap is particularly relev ant for 2 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T F igure 1: Multi-modal fusion at scale, left) data containing the featur es, and right) its r epr esentation in an embedding. applied domains such as climate adaptation, ecological monitoring, and infrastructure planning, where understanding the structure of model outputs can inform more targeted and ef ficient use of these representations. Furthermore, applied studies explicitly report limitations of GAEF in contexts where interpretability and time sensitivity matter (e.g., agriculture), pointing to low interpretability and other practical constraints such as spatial transferability and limited time sensitivity (Ma et al., 2025). Evidence suggests that GAEF content is rich but poorly characterized: its geometric properties (e.g., normalized embeddings on the unit hypersphere) and its use as a consistent global basis are known, but there is no operational consensus on which dimensions correspond to which variables/processes, nor on ho w to robustly v alidate these associations (Liu et al., 2025; Rahman, 2026). Recent studies hav e begun to sho w that GAEF embeddings contain physically meaningful information that extends well beyond their original mapping objecti ves. Rahman (2026) demonstrated that indi vidual embedding dimensions can be associated with continuous en vironmental v ariables — such as temperature, vegetation indices, hydrology , and terrain — and that these associations enable lar ge language models to generate interpretable descriptions of land surface conditions with high accuracy . Qu et al. (2026) sho wed that GAEF embeddings provide richer and more transferable representations of basin characteristics than hand-crafted hydrological attributes, impro ving streamflow prediction in ungauged basins by capturing inte grated en vironmental signals related to topography , ve getation, and soil properties. K och et al. (2025) further demonstrated that GAEF embeddings constitute effecti ve covariates for modelling soil organic carbon and water table depth in peatlands, performing comparably to expert-deri ved remote sensing inputs while offering substantially lo wer data preparation barriers. Howe ver , these approaches have focused primarily on associating embeddings with continuous environmental variables, optimizing their predictive utility for specific downstream tasks, or evaluating their performance against expert-deri ved co variates — without addressing a more fundamental question: whether the GAEF embedding space itself exhibits a hierarchical functional or ganization, in which individual dimensions play dif ferentiated roles ranging from high specialization in particular surface conditions to the encoding of broader , shared en vironmental gradients. Identifying this organizational structure is a prerequisite for systematic, task-independent interpretation of what embed- dings encode. Once such a functional chassis is characterized, it becomes possible to deriv e grounded interpretations of embedding dimensions in relation to the discrete categories that structure the Earth’ s surface. Among these, land cover stands out as the natural entry point: it is the most immediately perceptible e xpression of the Earth’ s surface, globally consistent, spatially pervasiv e, temporally persistent, and foundational to a wide range of Earth system processes — from carbon cycling and biodiv ersity to hydrology and climate regulation. Consequently , there is currently no methodological framework that allows the GAEF embedding space to be decomposed into distinct functional roles along a spectrum from specialization to generalization, which limits its use in scientific analysis and knowledge-based decision-making. T o address this gap, we propose a functional interpretability framework that analyzes the role of embedding dimensions in representing land cover or ganization. Specifically , we inv estigate whether the embedding space exhibits a structured, functional org anization and whether its dimensions can be systematically categorized according to their contribution to class discrimination and spatial interactions. Our central hypothesis is that embedding dimensions encode differentiated functional roles rather than homogeneous information, and that land co ver classes are characterized not 3 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T by individual dimensions in isolation, b ut by structured interactions among them. T o test this, we combine large-scale experimentation with structural analysis to characterize how embedding dimensions jointly relate to land cov er classes. The contributions of this work are fourfold: (i) a functional interpretability framework for geospatial embedding dimensions targeting global land cov er , (ii) empirical evidence of non-uniform functional beha vior across embedding dimensions, (iii) a taxonomy of specialist and generalist embedding dimensions along a functional spectrum, and (iv) a conceptual link between latent representations and geographic structures such as core regions and transition zones. 2 Materials and Methods 2.1 Overall study design This study proposes a functional interpretability framew ork for GAEF based on a combination of massiv e experimenta- tion and structural analysis of the embedding space. The approach is based on the use of embeddings generated by GAEF , which are high-dimensional representations that integrate spatial, temporal, and multi-modal information about the Earth’ s surface into a unified latent space (Brown et al., 2025). The proposed methodology is structured around two complementary components: a massi ve experimental exploration aimed at empirically characterizing the discriminativ e behavior of embeddings; and a structural analysis based on feature importance patterns and progressi ve ablation that allo ws interpretations of the functional or ganization of the representation space in terms of land cover types and their interactions. These components are visualized with an interactiv e Dashboard that enables exploratory analysis of the embedding space structure and its relationship to land cov er classes (Guthrie and Benavides, 2025b). This design seeks to transcend e valuation based e xclusiv ely on performance metrics, incorporating an interpreti ve perspectiv e that allows the embedding space to be broken do wn into differentiated semantic functions, thus addressing the interpretability limitations reported in GAEF applications (Figure 2). 2.2 Data and repr esentations The analysis is based on the integration of two main sources of information: land cover labels and the 64 GAEF embeddings. The labels come from the European Space Agenc y’ s (ESA) W orldCover 2020 product, which pro vides a global classification of the Earth’ s surface into 11 discrete categories (Zanaga et al., 2021). ESA W orldCov er was selected for sev eral reasons: its land cov er classes are globally generalizable, pro viding a consistent cate gorical framew ork across div erse geographic contexts; its high spatial resolution (10 m) supports fine-grained discrimination between classes; and its temporal alignment with the period over which GAEF was constructed ensures that the land cover labels correspond to the same conditions encoded in the embeddings. Crucially , ESA W orldCov er is independent from the training data used in the construction of GAEF embeddings, meaning that the classification experiments e valuate the embeddings ag ainst an external reference rather than recov ering labels already encoded during training. These classes represent the categorical structure of the Earth’ s surface and constitute the target v ariable in the classification experiments. The explanatory v ariables used are the GAEF embeddings, which correspond to the 64-dimensional vectors produced from the integration of multiple observation sources, including optical sensors, radar , climate variables, and other Earth observation products (Bro wn et al., 2025; Houriez et al., 2025). 2.3 Massive experimental exploration of embeddings The first methodological component consists of a large-scale automated experiment designed to e valuate the discrimina- tory capacity of embeddings in land co ver classification tasks. This experiment in volv es the execution of more than 130,000 independent analyses, each structured as a binary classification in which a target land co ver class is e valuated against an aggre gation of all remaining 10 classes, thereby isolating the embedding dimensions most informati ve for distinguishing each indi vidual land co ver type. T o explore the fundamentals of experimental design and run experiments independently , refer to the GAEF Land Cov er Classification App and to the Python pipeline de veloped to run the full set of experiments (Bena vides, 2025, 2026). T o ensure that each experiment contains suf ficient representation of the target class, a class-presence-guided spatial selection procedure is implemented using Google Earth Engine. For each analysis, a continent is randomly selected from a global set and the ESA W orldCov er layer is queried to identify a location where the tar get land cover class is present. Then a rectangular region of interest (R OI) of randomized dimensions ( 0 . 1 − 1 . 0 ) is constructed around this 4 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T F igure 2: Flowchart of the methodological framework: fr om embeddings to functional semantics. The pipeline inte grates (1) ESA W orldCover 2020 cate gorical labels and AEF 64-dimensional embeddings as inputs, (2) a massive experimental fr amework executing mor e than 130,000 binary classification experiments using Random F orest, Gr adient Boosting, XGBoost, and LightGBM with pr ogr essive feature ablation acr oss 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. location. This approach ensures that each R OI contains pixels of the target class while preserving spatial variability across experiments so that the global di versity of land co ver conditions is represented. In cases where no target class pixels are found within the selected continent, the system falls back to a random R OI within the continental bounds. Once the analysis re gion has been defined, the land co ver labels and embedding dimensions corresponding to the selected pixels are extracted simultaneously via stratified sampling, ensuring balanced representation of the target class and all other classes. These data are used to b uild training and v alidation sets, in a split of 75 / 25 , allo wing the discriminativ e ability of each embedding to be e valuated between classes. For each experiment, one of four machine learning algorithms is randomly selected: Random Forest, Scikit-Learn’ s Gradient Boosted Trees, XGBoost, and LightGBM. All four are suitable for capturing nonlinear relationships and have proven effecti ve in remote sensing classification tasks. Each experiment follo ws a training and e valuation scheme in which a model is first trained using all 64 embedding dimensions, and the relati ve contrib ution of each dimension to the classification is estimated using the nati ve feature importance scores of the trained model, based on Mean Decrease in Impurity (MDI) across decision tree splits. Based on this ranking, a progressiv e ablation procedure is implemented in which models are sequentially retrained using subsets of the embedding dimensions, incrementally expanding from the single most important dimension up to the top 30, as ordered by their MDI-based importance scores. This approach allows for the analysis of the redundancy and complementarity of the information contained in the embeddings, as well as the e valuation of ho w many dimensions are necessary to achieve adequate performance. For each land cover class, the tipping point is defined as the minimum number of dimensions at which mean classification performance across all e xperiments reaches 98% of the baseline performance obtained with all 64 dimensions (Figure 5). The dimensions ranked within this tipping point constitute the minimum subset for that class. 5 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T The results of each experiment — including performance metrics, variable importance, geographic location, and algorithm used — are systematically recorded in a structured database. This database constitutes the main input for subsequent interpretiv e analysis, providing lar ge-scale empirical evidence on the beha vior of these embeddings. 2.4 Structural analysis of the embedding space The second methodological component focuses on interpreting the embedding space based on the experimental results. While the pre vious phase allowed us to identify the rele vance of each dimension in specific tasks, this phase seeks to characterize the ov erall organization of the representation space in relation to land co ver types. T o do this, we construct an association matrix that quantifies the relationship between each embedding dimension and the different land cov er classes (T able 1). For each land cover class and each embedding dimension, the association score is computed as the normalized frequency with which that dimension ranked among the two most important features by MDI across all experiments targeting that class. Formally , the association score for a given class–dimension pair is defined as the number of experiments in which the dimension appeared among the top two MDI-ranked features, divided by the total number of experiments for that class, yielding a v alue between 0 and 1. This normalization ensures comparability across classes with different numbers of experiments. The resulting matrix provides a quantitative summary of the discriminative rele vance of each embedding dimension for each land cover class, and serves as the basis for the fingerprint plot (Figure 3), the conceptual embedding space representation (Figure 4), and the embedding univ erse visualization (Figure 6) implemented in the interactiv e Dashboard (Guthrie and Benavides, 2025b). T able 1: Excerpt of the association matrix quantifying the normalized frequency with which eac h embedding dimension ranked among the two most important featur es by MDI acr oss all experiments tar geting each land co ver class. V alues rang e from 0 to 1, wher e higher values indicate str onger discriminative rele vance of that dimension for the corr esponding class impA01 impA02 impA03 impA04 · · · impA64 Bare/sparse ve getation 0.0267 0.0223 0.0350 0.0167 · · · 0.0409 Built-up 0.0044 0.1092 0.0478 0.0057 · · · 0.0040 Cropland 0.0163 0.0266 0.0564 0.0369 · · · 0.0179 Grassland 0.0211 0.0315 0.0552 0.0270 · · · 0.0161 Herbaceous wetland 0.0177 0.0449 0.0269 0.1090 · · · 0.0199 Mangrov es 0.0446 0.0103 0.0112 0.0162 · · · 0.0093 Moss/lichen 0.0093 0.0265 0.0338 0.0136 · · · 0.0285 Shrubland 0.0140 0.0278 0.0892 0.0178 · · · 0.0190 Snow/ice 0.0043 0.0041 0.1583 0.0778 · · · 0.0063 T ree cover 0.0314 0.0208 0.0454 0.0248 · · · 0.0315 Permanent water bodies ("W ater") 0.0381 0.0983 0.0659 0.0180 · · · 0.1386 T o classify embedding dimensions into functional categories, we lev erage the progressiv e ablation procedure described in Section 2.3. For each land co ver class, the progressi ve ablation procedure identifies the minimum subset of dimensions required to recover 98% of baseline classification performance. This threshold was selected because, at this le vel of performance recovery , the remaining 2% difference falls within the inherent uncertainty of the ESA W orldCov er 2020 reference labels themselves, meaning that further gains in classification performance would not meaningfully improv e agreement with the ground truth be yond what the reference product’ s o wn mapping accuracy permits. A dimension is considered “associated” with a land cov er class if it appears within that class’ s minimum subset. The number of classes with which an embedding is associated then determines its functional classification: dimensions appearing in only one class’ s minimum subset are interpreted as specialists , while those appearing in the minimum subsets of two, three, or four or more classes are classified as low- , mid- , and high-gener alists respectiv ely . For con venience, all non-specialist dimensions are collecti vely referred to as shar ed dimensions . Dimensions that do not appear in the minimum subset of any land cover class remain uninterpreted under this framew ork and are discussed in Section 4.6. This classification approach grounds the functional taxonomy in demonstrated classification performance rather than an arbitrary threshold on importance values, ensuring that each dimension’ s role is defined by its empirical contribution 6 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T to accurate land cover discrimination. This approach is supported by evidence that GAEF embeddings capture complex physical and en vironmental patterns, integrating multiple sources of information into a coherent representation (Brown et al., 2025), and that these representations can reflect relationships with en vironmental variables at different scales (Rahman, 2026). Additionally , Figure 3 presents what we term the “Embedding Fingerprint” plot — a visualization whose structure deliberately echoes that of a genomic fingerprint, where the presence and absence of band patterns at specific loci encode the identity of a biological sample. In our case, the on/off pattern of exclusi ve (blue) and shared (pink) embedding dimensions across positions A01–A64 encodes the functional identity of each land co ver class, much as specific band combinations in a DNA fingerprint encode the identity of an organism. Just as no two species share an identical banding pattern, each land cover class exhibits a distincti ve dimensional signature: specialist dimensions act as unique genetic markers exclusi ve to a single class, while shared dimensions — encompassing low-, mid-, and high-generalists — function as conserved sequences that appear across multiple classes, reflecting structural or spectral similarities between them. This visualization provides a compact summary of the dimensional composition required to achieve 98% classification accuracy for each land co ver type, highlighting differences in the balance between specialist and shared dimensions across classes and offering an intuitiv e entry point into the functional organization of the GAEF embedding space. 2.5 V isualization and interpretation of the uni verse of embeddings T o facilitate interpretation of the embedding space structure, we dev eloped the “What on Earth is AlphaEarth?” interactiv e dashboard, which visualizes the relationships between embedding dimensions and land cover classes in a unified graphical en vironment (Guthrie and Benavides, 2025b). The dashboard integrates classification outputs, embedding importance metrics, and geographic conte xt, enabling users to explore the functional or ganization of the embedding space (Figures 4 and 6). In its primary view , land cover classes are or ganized as main nodes arranged in a circular space, while embedding dimensions are represented as entities associated with these classes. Embedding dimensions classified as specialists are visualized as elements close to a single coverage, reflecting their exclusi ve association. The darkness of each embedding dimension represents the strength of its association with a particular land cov er class, where a darker green indicates a stronger association and a lighter green indicates a weaker one. In contrast, shared embedding dimensions — those associated with multiple land cover classes — are located in intermediate positions between their associated co verages, reflecting their role in capturing characteristics common to more than one class. Notably , such shared dimensions may also encode information relev ant to ecological transition zones, where land cover types co-occur with one another . This spatial arrangement is calculated using geometric centroids, positioning each embedding dimension according to its relationships with the land cover classes it characterizes. The Dashboard allows intuiti ve exploration of the structure of the embedding space, identifying patterns of special- ization, clustering, and connectivity between land cov er classes. This approach facilitates the interpretation of complex results and provides visual e vidence of the functional organization of the embeddings. The Dashboard comprises four integrated views. The Overview provides a synthesis of the global experiments, including model performance comparisons across the four algorithms employed and a global map of experiment locations. The Class Analysis vie w houses the “Embedding Univ erse” visualization (Section 3.6), class performance matrices (Section 3.2), embedding importance charts that enable direct comparison of dimensional contributions across classes (Sections 3.2, 3.3). The Geographic view maps indi vidual experiment bounding boxes onto satellite imagery alongside spatially aggregated performance heatmaps, enabling identification of geographic regions where the embedding space performs strongest or weakest (Supplementary Figure 2). Finally , a Chat interface serves as a companion to the experimental notebook (Section 2.3), guiding users through the configuration of ne w experiments. 2.6 Methodological integration and hypothesis validation The interpretation of each embedding dimension follows a structured inferential process that differs between specialist and shared dimensions. Specialist dimensions are interpreted directly through their exclusiv e association with a single land cov er class: because they contrib ute discriminativ ely to only one class, their functional meaning is anchored to the distincti ve biophysical properties of that class — such as the permanent water absorption signature of Permanent W ater Bodies (embedding A64) or the artificial material reflectance of Built-up areas (embeddings A09 and A35). Shared dimensions, howe ver , require a more deliberate interpretive process. For each shared dimension, we first ask whether a conceptual connection exists between the associated land cover classes that is independent of their geographic distribution — that is, whether the classes share a common spectral, structural, phenological, or ecological property that could plausibly be captured by a single embedding dimension. If such a conceptual link emerges, it is 7 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T F igure 3: Embedding F ingerprint plot illustrating the functional dimensional signature of each land cover class at the 98% performance reco very threshold. Analogous to a genomic fingerprint, wher e the pr esence and absence of bands at specific loci encodes biological identity , each r ow r epresents the c haracteristic pattern of exclusive (blue) and shar ed (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 shar ed dimensions — encompassing low-, mid-, and high-generalists — contrib ute to the classification of two or more classes. Gray positions indicate dimensions not requir ed to achie ve the 98% thr eshold for that class. Land cover classes are or dered fr om top to bottom by incr easing spectral complexity , quantified as the mean pixel-wise variance acr oss Sentinel-2 spectral bands computed over 100 randomly selected imag es per class after standard atmospheric and r adiometric corr ections. Spectrally homo geneous classes such as Snow/ice and W ater — which e xhibit low inter-pixel variability due to their uniform reflectance signatur es — appear at the bottom, while spectrally heter ogeneous classes such as Built-up — whose radiometric r esponse varies substantially acr oss r ooftops, r oads, and impervious surfaces — appear at the top, r eflecting the incr easing number of embedding dimensions r equired to c haracterize them. adopted as the primary interpretation. When no immediate conceptual connection is apparent, we turn to geographic evidence, e xamining the spatial co-occurrence of the associated classes through maps, satellite imagery , and published literature to determine whether their association reflects a geographically grounded en vironmental relationship. This two-stage interpreti ve approach — prioritizing conceptual parsimony before geographic contingency — ensures that the functional meanings assigned to shared dimensions are grounded in either biophysical logic or empirically observable spatial patterns, rather than in arbitrary statistical co-occurrence. 3 Results 3.1 Empirical evidence of the inf ormational efficiency of embeddings The results of the lar ge-scale experimental exploration re veal that the discriminatory capacity of GAEF embeddings is not uniformly distributed in the representation space. A small subset of dimensions concentrates information necessary to differentiate land co ver types, and models trained with a limited number of prioritized dimensions achieve performance lev els comparable to those obtained with the full set of 64 embeddings (Figure 5a,b). 8 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T F igure 4: Conceptual structur e of the embedding space r epresentation, illustrating the planet-moon or ganization of the embedding universe. Blue circles r epr esent land cover classes (planets), while green cir cles r epr esent exclusive embedding dimensions (moons) whose color intensity encodes association strength — dark gr een indicating str ong discriminative r elevance, medium gr een indicating moderate r elevance, and light gr een indicating weak r elevance for that class. Gold diamonds repr esent shared embedding dimensions that connect multiple land cover classes, positioned at the geometric centr oid of their associated classes and reflecting ecotonal or spectrally ambiguous r egions wher e land cover boundaries overlap. This conceptual layout serves as the interpr etive basis for the embedding universe visualization implemented in the interactive dashboar d (Guthrie and Benavides, 2025b) and shown in F igure 6. Across the 11 ESA W orldCov er classes, the minimum number of embedding dimensions required to recover 98% of baseline classification performance v aries substantially (Figure 5c): W ater requires only two dimensions, Mangrov es four , and Built-up fiv e, while classes with less distinctiv e spectral or structural signatures — such as Moss/lichen and Shrubland — require up to 12 dimensions. In all cases, performance curves exhibit a characteristic plateau, with rapid gains from the first few dimensions followed by diminishing returns as additional dimensions are added (Supplementary Figure 1). This pattern suggests the existence of considerable redundanc y in the embedding space alongside a core of highly informati ve dimensions that capture distinctiv e characteristics of each land cover type. The variation in tipping points across classes further indicates that the embedding space encodes class-specific information at dif ferent lev els of concentration — classes with unique spectral or spatial signatures are captured by fe wer , more specialized dimensions, while classes sharing characteristics with others require a broader set. This finding is consistent with the nature of representations learned in foundation models, where information tends to be or ganized into distributed b ut non-homogeneous latent structures (Rahman, 2026). Beyond its theoretical implications, this result has direct practical relev ance: embedding prioritization enables substantial reductions in computational cost without meaningful loss in classification accuracy . Depending on the land cov er class, restricting inference to the minimum embedding subset reduces classification time by 20% to 80% relativ e to the full 64-dimension baseline, of fering a pathway tow ard more efficient deplo yment of foundation model embeddings in operational contexts. 9 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T (a) (b) (c) F igure 5: Pr ogr essive ablation curves and minimum embedding dimensions by land cover class. P anels (a) and (b) show eight classification metrics as a function of MDI-r anked embedding dimensions for W ater and Shrubland r espectively (r ed = mean acr oss experiments; blue dashed = 64-dimension baseline; gr een dotted = tipping point at 98% baseline r ecovery). W ater r equires as fe w as 2 dimensions while Shrubland r equir es up to 12, reflecting dif fer ences in spectral distinctiveness. P anel (c) summarizes the minimum dimensions r equir ed to achie ve 98% of baseline performance acr oss all 11 ESA W orldCover 2020 land cover classes. 10 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 3.2 Emergence of highly specialized embedding dimensions The analysis of the importance of variables reveals the recurring presence of embedding dimensions with a high contribution to the discrimination of specific co verages. These dimensions show a strong association with a single class, maintaining limited relev ance in other categories. This pattern suggests the existence of highly specialized embedding dimensions that capture distincti ve characteristics of certain land cover types. In particular, it is observed that land cover types with clearly distinguishable physical signatures, such as bodies of w ater or urban areas, tend to ha ve embedding dimensions with a high concentration of importance. The emergence of these specialized embedding dimensions is consistent with the ability of foundation models to encode physical and environmental properties in high-dimensional latent spaces (Rahman, 2026). Howe ver , this work does not seek to directly associate embedding dimensions with specific physical v ariables such as temperature or elev ation. Instead, it interprets their behavior functionally , examining how individual dimensions contribute to discriminating between land cov er classes and capturing spatial interactions among them. T able 2: Exclusive embedding dimensions and their associated ESA W orldCover 2020 land cover classes, listing dimensions that appear solely in one class’ s minimum subset alongside a biophysical interpr etation of the distinctive spectral, structur al, or phenological pr operties that justify their exclusive association. Exclusive Embedding Dimension Associated Land Cover Class Distinctive trait that can justify an exclusi ve embedding A09, A35 Built-up Urban 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, A50 Cropland Annual crops so wn and harvested within a 12-month cycle, with very re gular phenology (well-marked greenness peaks), geometric patterns (plots, furro ws) and often intensi ve management. Al- though they can be confused with grasslands, certain re gions of the feature space (e.g. very homogeneous or highly irrig ated crops) carry signatures specific to cropland, e xplaining their exclusi ve embeddings. A05, A27 Mangrov es Ever green woody ve getation tolerant to salinity in the intertidal zone. They sho w dense canopy lik e the Tree cover class, but with coastal linear patterns, water mixed between cro wns and specific NIR/SWIR signatures due to permanent waterlogging. This combi- nation (dense forest, brackish w ater and coastal context) is unique to mangrov es. A04, A11, A25, A29 Herbaceous wetland Areas dominated by herbaceous v egetation that is permanently or regularly flooded (fresh, brackish or saline water). The y combine moderate–high ND VI with very high water indices (ND WI, etc.) and floodplain textures; no other type combines herbaceous co ver and long-lasting flooding in this way . A18, A21, A26 Shrubland Zones dominated by natural shrubs ( ≥ 10% cov er), typical of semi- arid biomes. Although they mix with grassland and bare/sparse ve getation, subtypes where shrub height, density and spatial pat- tern (e.g. dense Mediterranean shrublands) generate signatures distinct enough to occupy an exclusi ve subspace. A16, A23 T ree cover Regions dominated by trees ( ≥ 10% cov er), including plantations, with very high NIR, strong shado ws and closed canopy textures. This 3-D structure and phenological persistence (especially in e ver- green forests) clearly separate them from shrubland and grassland, allowing embeddings to specialise in representing “dense forests”. Continued on next pa ge 11 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T T able 2 (continued fr om pr evious pag e) Exclusive Embedding Dimension Associated Land Cover Class Distinctive trait that can justify an exclusi ve embedding A41 Grassland Areas dominated by natural herbaceous v egetation (grasslands, sav annas, prairies) with ≥ 10% cov er and no dominant woody stems. They sho w marked seasonality b ut a homogeneous struc- ture, without long-lasting flooding (unlike Herbaceous wetland) or intense geometric traces (unlike Cropland). Embedding A41 likely captures “well-defined grassland” far from crops and shrubs. A44, A57 Moss/lichen Surfaces dominated by mosses and/or lichen, typical of tundra and cold or very arid rock y en vironments. Their very low structure and spectral behaviour (moderate ND VI, fine textural patterns o ver rock or bare soil) dif fer markedly from tall herbaceous vegetation and shrubs, justifying exclusi ve embeddings for these cryptog amic en vironments. A63 Bare/sparse ve getation Lands with soil, sand or e xposed rock and < 10% ve getation throughout the year . This is the “purest” bare soil signature: high reflectance in VIS and SWIR, minimal chlorophyll signal, and very little temporal v ariation. No other class maintains such consis- tently lo w vegetation co ver year -round, so a dedicated embedding is reasonable. A64 Permanent water bodies Permanent water bodies ( ≥ 9 months/year), with v ery strong absorption in NIR and SWIR, clear linear or polygonal geom- etry (ri vers, lakes) and lo w spectral v ariability . This near -zero NIR/SWIR reflectance and smooth te xture is unique compared to any other class, so a specific embedding for w ater is fully ex- pected. 3.3 Evidence of shared embedding dimensions between co verages Additionally , the analysis sho ws the existence of embedding dimensions that make significant contrib utions across multiple co verages. These dimensions are not exclusiv ely associated with one class, but rather participate in the discrimination of different types of cov erage. This behavior suggests that certain embedding dimensions capture common patterns or interactions between land cov er classes, which is particularly rele vant in conte xts where classes are not strictly separable. From a geographical perspectiv e, this type of representation is consistent with the existence of transition zones or ecotones, where characteristics of multiple land cov erages coexist. From a functional perspecti ve, it is also consistent with land co ver classes that share spectral, phenological, or structural characteristics, such as croplands and grasslands, which may exhibit similar seasonal greenness patterns despite differing in management and land use. The presence of these shared embeddings is consistent with the ability of GFMs to integrate multi-source information and capture complex spatial dynamics (Brown et al., 2025). In this sense, these dimensions cannot be interpreted as e xclusive attrib utes, but rather as components that reflect relationships between classes. T able 3: Shar ed embedding dimensions and their functional interpretation, gr ouped by generalist cate gory (low-, mid-, and high-gener alist) accor ding to the number of land co ver classes in whose minimum subset the y appear . F or eac h dimension, the associated ESA W orldCover 2020 classes and a common ecological–spectral trait ar e reported, inferr ed thr ough the two-stage interpr etive pr ocess described in Section 2.6. Embedding Dimension Associated ESA Land Covers Categorization Identifier Common trait (ecological–spectral synthe- sis) A07 Cropland, Moss/lichen Low-generalist Discontinuous ve getation Areas with lo w and discontinuous ve getation ov er mineral soil: croplands in fallo w or early phenological stages and lichen mats sho w a soil and sparse v egetation mixture with high visible albedo and marked seasonal v ariability . Continued on next pa ge 12 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T T able 3 (continued fr om pr evious pag e) Embedding Dimension Associated ESA Land Covers Categorization Identifier Common trait (ecological–spectral synthe- sis) A08 Built-up, Moss/lichen Low-generalist Sparse ve getation ov er bright surfaces Bright, sparsely ve getated substrates such as roofs, concrete and rock with lichens share low chlorophyll signal and rough te xtures that the model groups as mineral surf aces with minimal ve getation. A15 Moss/lichen, Snow/ice Low-generalist High latitude and high albedo High-latitude or high-altitude cold en viron- ments where sno w/ice and lichen mats ov er rock produce bright visible reflectance and spatially homogeneous cryogenic landscapes. A28 Grassland, T ree cov er Low-generalist Dense green ve ge- tation Landscapes dominated by green v egetation with strong NIR response where the main difference is canopy density; the embedding captures forest–grassland mosaics with inter - mediate canopy fraction. A34 Shrubland, T ree cover Low-generalist W oody perennial W oody perennial ve getation with relati vely high biomass; shrublands and forests dif fer in height but share canopy te xtures and similar spectral and SAR responses. A36 Bare/sparse, Herbaceous wetland Low-generalist Soil-moisture gra- dient T ransition from dry bare soil to wet herba- ceous substrates where e xposed mud and sea- sonal wetlands resemble moist bare surf aces, capturing gradients of soil moisture. A40 Grassland, Shrubland Low-generalist Open lo w–medium ve getation Continuum of open v egetation of lo w to medium height typical of grasslands and shrub- lands that share discontinuous structure and similar spectral behaviour . A43 Grassland, Moss/lichen Low-generalist Low-height vegeta- tion Surfaces with creeping or very lo w vegetation such as sparse grasslands or lichen-dominated tundras, producing moderate ND VI and fre- quent mixing with exposed soil. A53 Mangrov es, Moss/lichen Low-generalist Extreme conditions V egetation adapted to e xtreme environments: saline mangrov es and cold or arid lichens both sho w specialized perennial ve getation and complex te xtures linked to en vironmental stress. A55 Bare/sparse, Moss/lichen Low-generalist Rocky sparse v ege- tation Mineral substrates dominated by rock or soil with extremely sparse vegetation where lichen patches barely increase ND VI above bare soil lev els. A59 Bare/sparse, Moss/lichen Low-generalist High-reflectance sparse ve getation Reinforces the embedding subspace associated with high-reflectance surfaces with very lo w ve getation cover where lichens and bare soil are spectrally mixed. A10 Bare/sparse ve getation, Herbaceous wetland, Man- grov es Mid-generalist Soil–flooding gra- dient Gradient from dry bare soil to flooded herba- ceous wetlands and mangro ve systems, rep- resenting strong soil moisture and flooding dynamics across floodplains and coastal zones. A13 Bare/sparse ve getation, Grassland, Shrubland Mid-generalist Fractional ve geta- tion in drylands Semi-arid landscapes characterized by gradual transitions from bare soil to sparse grasslands and shrublands where v egetation fraction varies continuously . Continued on next pa ge 13 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T T able 3 (continued fr om pr evious pag e) Embedding Dimension Associated ESA Land Covers Categorization Identifier Common trait (ecological–spectral synthe- sis) A14 Cropland, Grassland, Herbaceous wetland Mid-generalist Herbaceous ve geta- tion systems Land cov ers dominated by herbaceous ve g- etation including crops, natural grasslands and wetlands that share strong seasonality and similar canopy spectral signatures. A37 Bare/sparse ve getation, Built-up, Moss/lichen Mid-generalist Highly reflecti ve sparse ve getation Reflectiv e mineral or artificial substrates such as rock, bare soil and urban surfaces with mini- mal ve getation signals and strong brightness in visible bands. A51 Cropland, Shrubland, T ree cover Mid-generalist Agroforestry mo- saics Agricultural landscapes where crops coe xist with woody v egetation such as agroforestry systems, plantations or silvopastoral mosaics. A52 Bare/sparse ve getation, Shrubland, T ree cover Mid-generalist W oody ve getation gradient Gradual increase of woody vegetation from nearly bare surfaces to shrublands and open forests representing partial woody cover land- scapes. A56 Built-up, Moss/lichen, Snow/ice Mid-generalist High albedo sur - faces Surfaces with v ery high albedo and low chloro- phyll signal such as sno w , bright urban materi- als and rocks with lichens. A60 Cropland, Grassland, Shrubland Mid-generalist Cropland, grass- land, shrubland axis Rural matrices where agricultural fields, f al- lows and semi-natural ve getation coexist form- ing a strong continuum of herbaceous and woody land cov ers. A62 Grassland, Shrubland, Permanent water bodies Mid-generalist Riparian mosaics Landscapes where open v egetation interacts with water bodies such as riparian corridors and piedmont areas combining v egetation reflectance and water absorption signatures. A03 Cropland, Grassland, Shrubland, Snow/ice High-generalist Seasonal snow o ver open ve getation T emperate and cold landscapes where open herbaceous ve getation periodically becomes snow covered, generating strong seasonal spectral changes. A45 Bare/sparse ve getation, Grassland, Shrubland, Snow/ice High-generalist Arid–semiarid en vironments Gradients of sparse v egetation typical of arid or semi-arid regions occasionally influenced by seasonal snow e vents. A61 Bare/sparse ve getation, Cropland, Moss/lichen, Snow/ice High-generalist High latitude and mountain mosaics Highly heterogeneous landscapes typical of high latitudes or mountains where bare soil, marginal croplands, lichens and snow coexist within extremely seasonal en vironments. 3.4 High-generalist embedding dimensions and global behavior Among the shared dimensions reported in T able 3, high-generalist dimensions (those associated with four or more land cover classes) are distinguished by their consistently low contribution to individual classification tasks. These dimensions have a more homogeneous distribution of importance across coverages and do not show a dominant association with any class. This behavior suggests that these dimensions capture more general patterns, possibly associated with large-scale en vironmental gradients or v ariables that are not decisive for the discrimination of specific land co ver types. This type of representation is consistent with the global nature of GAEF embeddings, which seek to provide a consistent description of the state of the Earth’ s surface on a planetary scale (Houriez et al., 2025). Although these dimensions do not contribute significantly to indi vidual class discrimination, their presence suggests that robust land cov er classification depends on the interplay between specialist and shared dimensions, not specialist dimensions alone. 14 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 3.5 Functional organization of the embedding space The integration of the observed patterns reveals a functional structure in the embedding space, characterized by the coexistence of dif ferent types of dimensions according to their beha vior in classification tasks. The results suggest a spectrum of functional roles: from specialist dimensions with high discriminatory capacity for a single land cover class, through low- and mid-generalist dimensions that capture relationships between tw o or three classes, to high-generalist dimensions that reflect broader en vironmental patterns. This org anization emerges consistently from empirical analysis, rather than from any e xplicit structural constraint known to the authors, suggesting that the embedding space encodes structured information related to the geographical organization of the territory . This finding contributes to advancing our understanding of foundation models, by showing that their representations are not arbitrary , but rather reflect patterns that can be interpreted from an ecological and spatial perspectiv e. 3.6 V isual repr esentation of the universe of embedding dimensions The embedding universe visualization (Section 2.5) provides an integrated vie w of the identified patterns. Figure 6, generated using the interacti ve Dashboard (Guthrie and Bena vides, 2025b), illustrates the embedding space at a 98% baseline performance recovery threshold. At this threshold, a predominance of specialist embedding dimensions is observed, with most dimensions positioned in close proximity to a single land cover class, suggesting a high capacity for specialization in the representation space. Shared dimensions, though fewer in number , occupy intermediate positions that highlight inter-class relationships. This visualization provides qualitativ e evidence of the functional or ganization described in the preceding analyses. F igure 6: “Embedding Universe” visualization at a 98% baseline performance r ecovery thr eshold, generated using the interactive Dashboar d (Guthrie and Benavides, 2025b). Green nodes r epr esent specialist dimensions (darker = str onger association); gold nodes repr esent shar ed dimensions. See Section 2.5 for a full description of the visual encoding. T o interact with this plot, access this link. 15 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 3.7 Summary of findings Overall, the results obtained do not seek to conclusiv ely validate a hypothesis or to attribute definiti ve and unique interpretations to individual embedding dimensions, b ut rather to provide empirical evidence in support of a functional interpretability frame work. The observation of consistent patterns in the importance of embeddings, their dif ferential behavior between land cov er classes, and their or ganization in the representation space suggests that it is possible to characterize these dimensions in functional terms — while ackno wledging that the interpretations proposed here are task-dependent, biophysically informed approximations rather than fixed semantic labels. These findings moti vate the conceptualization of the embedding space as a structured system, in which dimensions can be interpreted according to their role in representing land cover and its interactions. In this sense, the results constitute the empirical basis for the proposed interpretability framework, which rev eals a hierarchical functional structure within the embedding space — one that, once characterized, provides a transferable analytical chassis through which embeddings can be interpreted in relation to an y geospatial domain or dataset, be yond the land cover context explored here. 4 Discussion 4.1 Interpr etability in GFMs: from physical variables to functional structur es GFMs ha ve demonstrated significant progress in Earth observ ation tasks, thanks to their ability to inte grate multi-source data into latent representations of high dimensionality (Brown et al., 2025; Houriez et al., 2025). Ho wever , this progress has been accompanied by a fundamental challenge: the lack of interpretability of embeddings, which encode en vironmental information into abstract v ectors with no direct correspondence to observable physical variables (Rahman, 2026). Existing approaches hav e addressed this limitation primarily by associating embedding dimensions with continuous variables, such as temperature, precipitation, or ele vation, seeking to identify explicit physical relationships within the latent space (Rahman, 2026). While these approaches represent an important advance, they focus on a physical-v ariable interpretation, which does not necessarily capture the categorical or ganization of the territory in terms of land cov er . In contrast, this paper proposes a complementary approach based on functional interpretability , in which embeddings are interpreted according to their role in discriminating between coverages and representing their interactions. This approach allo ws the latent space to be connected to fundamental geographic concepts, such as coverage units and transition zones, extending the scope of interpretability be yond traditional physical variables. 4.2 Functional organization of the embedding space The results suggest that the embedding space has a non-uniform structure, where dimensions play different roles. This org anization can be interpreted as a functional spectrum, from specialist dimensions that capture the distinctiv e characteristics of a single land cover class, through low- and mid-generalist dimensions that encode relationships between two or three classes, to high-generalist dimensions that reflect broader en vironmental patterns. Notably , specialist dimensions, which concentrate importance on unique spectral-structural signatures like “W ater” or “Built-up”, effecti vely represent the “core” of land-cover units. Meanwhile, the identification of mid-generalist dimensions suggests that the embedding space captures the fuzzy logic of environmental transitions. These mid- generalist dimensions act as semantic bridges, representing ecotones where classes o verlap — such as the moisture- driv en gradient between ’Bare soil’ and ’Herbaceous wetlands’. This functional organization implies that foundation models do not merely memorize pixels b ut encode a structured blueprint of en vironmental perception. This functional structure emerges consistently from empirical analysis, suggesting that embeddings are not arbitrary representations but rather reflect regularities inherent in the or ganization of the territory . In this sense, the embedding space can be understood as a system in which dif ferent dimensions capture dif ferent le vels of abstraction of geographical reality . This finding is consistent with evidence that foundation models are capable of learning rich and generalizable representations from heterogeneous data (Brown et al., 2025), but it offers a new perspectiv e by suggesting that these representations also possess a structure that can be interpreted in terms of discrete spatial categories and their interactions. 16 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 4.3 Specialist embeddings and coverage cor e repr esentation The identification of embeddings with high specialization in certain co verages suggests that the latent space contains dimensions that capture distincti ve characteristics of specific categories of the territory . This behavior is particularly evident in co verages with well-defined physical signatures, where the separability between classes is clearer . From an interpretativ e perspectiv e, these embeddings can be understood as representations of the “cores” of land cov er classes, i.e., regions where the characteristics of a class are dominant and relati vely homogeneous. In these contexts, specialized embeddings enable accurate discrimination, acting as robust indicators of the presence of a land cov er class. This type of representation is consistent with the ability of foundation models to capture relev ant physical properties of the en vironment, although in this case the interpretation is performed in terms of geographic categories rather than continuous variables (Rahman, 2026). 4.4 Low- and mid-generalist dimensions and ecotone r epresentation One of the most significant contributions of this work is the identification of generalist embedding dimensions associated with a limited number of land cover classes. These dimensions serve a dual role: they capture characteristics common to classes that ov erlap spatially or functionally , as in transition zones or ecotones where coverages lack clearly defined boundaries, and they complement specialist dimensions in the classification of indi vidual land cov er classes where no single dimension is sufficient on its o wn. The existence of these dimensions suggests that foundation models not only capture discrete patterns, but also the interactions between them, integrating information from multiple sources and spatial scales. This behavior is consistent with the nature of en vironmental systems, where transitions between land covers are frequent and play a key role in the dynamics of the territory . Their association with tw o or three specific classes makes low- and mid-generalist dimensions particularly suited to representing the inter-class relationships characteristic of ecotonal boundaries. In this sense, these dimensions constitute a bridge between the continuous representation of latent space and the discrete structure of coverages, allo wing the complexity inherent in geographic systems to be captured. This aspect has been little explored in the literature, despite its rele vance for applications such as monitoring land use changes or detecting degradation processes. 4.5 High-generalist dimensions and en vironmental gradient r epresentation In addition to specialist, low-, and mid-generalist dimensions, high-generalist dimensions are identified that exhibit more general behavior , without a dominant association with any individual land co ver class. These dimensions can be interpreted as representations of large-scale en vironmental gradients, such as climatic or topographic conditions. This type of dimension reflects the global nature of GFMs, which seek to provide a consistent representation of the planet across different regions and contexts (Houriez et al., 2025). Although these dimensions are not highly discriminati ve in classification tasks, their presence indicates that the embedding space incorporates information relev ant to the general characterization of the en vironment. From an interpretive perspective, high-generalist dimensions can be considered as a higher lev el of abstraction, capturing properties of the system that transcend the discrete categories of land co ver . 4.6 Uninterpr eted dimensions and their significance Of the 64 embedding dimensions in GAEF’ s latent vector space, the preceding analyses yield functional interpretations for 43. The remaining 21 dimensions were not found within the minimum embedding subset for any land co ver class across more than 130,000 independent experiments. These uninterpreted dimensions may encode information outside the land cov er domain — such as broader en vironmental gradients, atmospheric conditions, or properties of the b uilt en vironment — or may represent aspects of the Earth’ s surface that the current experimental frame work, focused on land cov er classification, is not designed to capture. This constitutes an important direction for future research. Cross-domain experiments tar geting alternativ e domains (e.g., infrastructure, climate, or land use) could rev eal whether these dimensions carry structured information that becomes legible under a different contextual lens. Additionally , targeted validation experiments using other datasets may help determine whether these dimensions truly encode subtle en vironmental signals that fall belo w the discrimination threshold in the current design. 17 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 4.7 Implications for inter pretability and r emote sensing The proposed frame work has important implications for the interpretability of foundation models in remote sensing. First, it allows interpretation to shift from a variable-based approach to a function-based approach, which is more suitable for analyzing complex geographic phenomena. Secondly , identifying different types of embedding dimensions opens up the possibility of designing more ef ficient and explainable models by selecting subsets of dimensions according to their function. For example, specialist dimensions could be used to impro ve accurac y in class-specific identification, while lo w- and mid-generalist dimensions could be key to impro ving change detection or edge delimitation. Thirdly , this approach allo ws the interpretation of models to be integrated with concepts specific to geography and ecology , facilitating their use in scientific and decision-making contexts. 4.8 Practical implications The findings of this study transcend theoretical interpretability , of fering immediate practical benefits for the deployment of GFMs in large-scale operational en vironments: • Computational Cost Reduction: By identifying that accurate land-cover classification ( ≥ 98% of baseline) requires as fe w as 2 to 12 dimensions, organizations can implement a dimension pruning strategy . This reduces the data footprint of high-resolution (10 m) global mosaics by up to 90%, significantly lowering storage costs and memory ov erhead during inference. • T ar geted Feature Selection: Instead of utilizing the full 64-dimensional embedding as a “black box, ” practition- ers can now select specialist dimensions tailored to specific domains . For instance, urban monitoring tasks can prioritize dimensions A09 and A35, while coastal management projects can focus on mangro ve-specific dimensions like A05 and A27. • Explainable AI (XAI) for Decision Support: The functional taxonomy provides a “semantic bridge” for non-expert stak eholders. By utilizing the “Embedding Uni verse” visualization, model predictions mov e from abstract vectors to justifiable geographic associations, facilitating trust in AI-dri ven climate adaptation and infrastructure planning. • Enhanced Change Detection: Identifying shared dimensions that represent ecotones allows for more sensiti ve monitoring of ecological transition zones. These dimensions can be used as early-warning indicators for land degradation or deforestation, where subtle shifts in the latent space signal transitions before they are e vident in categorical labels. 4.9 Limitations of the study Despite the contributions presented, this work has se veral limitations that should be considered. First, the analysis is based on importance measures deriv ed from machine learning models, which may be influenced by the algorithm used and the distribution of the data. This implies that the interpretation of the embeddings depends, to some extent, on the experimental conte xt. Secondly , the functional classification of embedding dimensions is deriv ed from progressiv e ablation at a 98% baseline performance recov ery threshold. While this threshold is grounded in classification performance rather than arbitrary cutoffs, different recovery thresholds would yield different minimum subsets and potentially different functional classifications. The sensitivity of results to this choice warrants further analysis. Additionally , although the proposed visualization facilitates interpretation, it represents a simplification of the embedding space, which is actually high-dimensional. Therefore, the relationships observed should be understood as an approximation of the actual structure of the system. Furthermore, although this study captures regions of interest distributed globally , the functional roles identified for each embedding dimension may shift when the frame work is applied at a re gional scale. Classification accuracy itself varies geographically (Supplementary Figure 2), suggesting that regional land cov er conditions influence the discriminatory capacity of the embedding space. In geographically constrained contexts, reduced land co ver di versity and en vironmental variability could diminish the discriminatory capacity of dimensions that are informativ e at a global lev el, effecti vely flattening the importance distribution. Finally , this study does not address the validation of the proposed interpretations in terms of specific physical variables or en vironmental processes, which constitutes a line of future work. 18 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 4.10 Future lines of r esearch The results obtained open up multiple lines of research. One relev ant direction is the validation of the proposed framew ork by associating embeddings with specific physical variables, which would allo w the integration of functional and physical interpretability approaches. Another line of research consists of analyzing the stability of the functional classification of embeddings in different regions and temporal scales, e valuating their capacity for generalization. Likewise, the study of the temporal dynamics of embeddings could provide information on processes of change in the territory . A third line of research rev olves around geographic variation. The geographic variation in classification performance rev ealed across the global experiments (Supplementary Figure 2) suggests that the functional roles of individual embedding dimensions may not be uniform across biomes or climatic zones. A systematic in vestigation of the geographic determinants of embedding discriminatory capacity constitutes a promising direction for future research. Additionally , the proposed frame work could be e xtended to other remote sensing tasks, such as change detection, semantic segmentation, or en vironmental variable prediction, exploring the role of different types of embeddings in each context. Finally , integrating these results into interactiv e analysis and decision-making systems — such as climate adaptation planning tools and en vironmental monitoring platforms — represents an opportunity to bring foundation models closer to practical applications, facilitating their interpretation by non-specialist users. 5 Conclusion This paper presents a functional interpretability frame work for GFMs, aimed at understanding the internal structure of embeddings in relation to the organization of territory . Based on an exploratory analysis in volving massiv e experimentation and structural characterization of the embedding space, the e vidence suggests that these representations are not arbitrary , but rather present consistent patterns that allow their beha vior to be interpreted in terms of land cover and their interactions. The results suggest that embedding space can be understood as a functionally organized system in which dimensions with dif ferent roles coexist: specialist dimensions highly specialized in representing specific land co ver classes, lo w- and mid-generalist dimensions associated with relationships between a limited number of classes, and high-generalist dimensions related to broader en vironmental gradients. This hierarchical organization allo ws the latent space of the foundation models to be connected to fundamental geographic concepts, such as land cover units and ecotones, offering an alternativ e perspectiv e to interpretability approaches based exclusi vely on physical v ariables. In this sense, the main contrib ution of this w ork lies in proposing an interpretability approach that shifts the focus from identifying variables to understanding functions within the embedding space. This approach opens up new possibilities for the analysis of foundation models in remote sensing, facilitating their use in scientific contexts and in informed decision-making. Finally , this study lays the groundw ork for future research aimed at v alidating and e xtending the proposed framew ork, including integration with biophysical v ariables, spatial and temporal stability analysis, and its application in different Earth observ ation tasks. T ogether , these adv ances will help bridge the g ap between the high predicti ve performance of foundation models and their scientific interpretation, promoting a more transparent and explainable use of these technologies. 6 Statements and Declarations 6.1 Competing Interests The authors hav e no competing interests to declare that are relev ant to the content of this article. 6.2 Clinical trial number Not applicable. 6.3 Funding The authors did not receiv e support from any or ganization for the submitted work. 19 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T 6.4 Ethical responsibilities of A uthors All authors hav e read, understood, and have complied as applicable with the statement on “Ethical Responsibilities of Authors” as found in the Instructions for Authors. 6.5 Data A vailability Statement No ne w primary data were created in this study . The analysis was conducted using publicly av ailable remote sensing data from the Copernicus Open Access Hub . The deriv ed data products supporting the findings of this study are av ailable from the corresponding author upon request. 6.6 Code A vailability Statement The source code for the “What on Earth is AlphaEarth?” interactive Dashboard is a v ailable at Guthrie and Benavides (2025a). The experimental notebook used to conduct the classification experiments is a vailable at Bena vides (2026). 7 Acknowledgments • I.F .B would lik e to thank his colleagues at the Roux Institute (The Institute for Experiential Artificial Intel- ligence and the Artificial Intelligence for Climate and Sustainability Groups) as well as the Gulf of Maine Research Institute. • J.G. would like to thank his f amily; his colleagues at the Sustainability and Data Sciences Lab and at Enodia, Inc.; and the Artificial Intelligence for Climate and Sustainability Group and the Institute for Experiential Artificial Intelligence for their encouragement and support. • J.E.A. would like to thank his colleagues at Uni versidad Católica de Manizales. • Y .A.G. would lik e to thank to Univ ersidad Católica de Manizales. • A.I.G. would like to thank her colleagues at Uni versidad Nacional de Colombia, Palmira Campus. • C.V .P . would like to thank his colleagues at Uni versidad Nacional de Colombia, Palmira Campus. 20 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T References Benavides, I. F . (2025). Alpha-Earth-Land-Cover -Classifier (V ersion v1) [Software]. Benavides, I. F . (2026). AlphaEarth-Interpretability-Experiments. Bodnar , C., Bruinsma, W . P ., Lucic, A., Stanley , M., Allen, A., Brandstetter , J., Garv an, P ., Riechert, M., W eyn, J. A., Dong, H., Gupta, J. K., Thambiratnam, K., Archibald, A. T ., W u, C.-C., Heider , E., W elling, M., T urner , R. E., and Perdikaris, P . (2025). A foundation model for the Earth system. Nature , 641:1180–1187. Bro wn, C. F ., Kazmierski, M. R., Pasquarella, V . J., Rucklidge, W . J., Samsikov a, M., Zhang, C., Shelhamer, E., Lahera, E., Wiles, O., Ilyushchenko, S., Gorelick, N., Zhang, L. L., Alj, S., Schechter, E., Askay , S., Guinan, O., Moore, R., Boukouv alas, A., and Kohli, P . (2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv pr eprint . Editorial (2025). T o wards responsible geospatial foundation models. Natur e Machine Intelligence , 7:1189–1189. Google (2025). AlphaEarth Foundations helps map our planet in unprecedented detail. Google DeepMind . Guthrie, J. and Benavides, I. F . (2025a). AlphaEarth_V iz. Guthrie, J. and Benavides, I. F . (2025b). “What on Earth is AlphaEarth?” Interactive Dashboard (V ersion v1) [Software]. Houriez, L., Pilarski, S., V ahedi, B., Ahmadalipour, A., Scully , T . H., Aflitto, N., Andre, D., Jaffe, C., W edner , M., Mazzola, R., Jeffery , J., Messinger, B., McGinley-Smith, S., and Russell, S. (2025). Scalable Geospatial Data Generation Using AlphaEarth Foundations Model. arXiv preprint . K och, J., Denager , T ., Stisen, S., Grev e, M. H., Møller, A. B., and Beucher , A. M. (2025). Modelling Carbon and Groundwater in Peatlands using AlphaEarth Embeddings. arXiv preprint . Liu, J., Qin, Q., Dong, G., W ang, X., Feng, J., Zeng, Z., and Cheng, T . (2025). Beyond AlphaEarth: T o ward Human-Centered Spatial Representation via POI-Guided Contrastiv e Learning. arXiv pr eprint . Ma, Y ., Shen, Y ., Swatantran, A., and Lobell, D. B. (2025). Harvesting AlphaEarth: Benchmarking the Geospatial Foundation Model for Agricultural Do wnstream T asks. ChatP aper . Qu, P ., Ouyang, W ., Zhang, C., Chai, Y ., Xu, S., Y e, L., Piao, Y ., Zhang, M., and Lu, H. (2026). Utilizing Earth Foundation Models to Enhance the Simulation Performance of Hydrological Models with AlphaEarth Embeddings. arXiv pr eprint . Rahman, M. (2026). Physically Interpretable AlphaEarth Foundation Model Embeddings Enable LLM-Based Land Surface Intelligence. arXiv preprint . Seydi, S. T . (2025). Deep Learning-Based Burned Area Mapping Using Bi-T emporal Siamese Networks and AlphaEarth Foundation Datasets. arXiv preprint . Zanaga, D., Kercho ve, R. V . D., Keersmaecker , W . D., Souverijns, N., Brockmann, C., Quast, R., W ev ers, J., Grosu, A., Paccini, A., V ergnaud, S., Cartus, O., Santoro, M., Fritz, S., Geor giev a, I., Lesi v , M., Carter , S., Herold, M., Li, L., Tsendbazar , N., Ramoino, F ., and Arino, O. (2021). W orldCover. ESA W orldCover 10 m 2020 v100 . 21 W H AT O N E A RT H I S A L P H A E A RT H ? A P R E P R I N T Supplementary Inf ormation Supplementary Figure 1: Embedding frequency distributions by land cov er class. F or each class, embedding dimensions ar e ranked by fr equency of appearance in the top 2 most important positions acr oss 10,000 experimental runs. Green bars indicate dimensions within the tipping point threshold; blue bars indicate remaining dimensions. Generated using the inter active Dashboar d (Guthrie and Benavides, 2025b). Supplementary Figure 2: Grid-based geographic performance heatmap of classification accuracy across global experiments. Eac h cell r epr esents the mean classification accuracy aggr e gated acr oss all experiments conducted within that geogr aphic grid cell. Red cells indicate higher accuracy (>90%), orange cells indicate moder ate accuracy (80–90%), and blue cells indicate lower accur acy (<80%). Generated using the interactive Dashboar d (Guthrie and Benavides, 2025b). 22