PaperlayerView exact revision on arXiv

Source-linked AI summary

Physically Interpretable AlphaEarth Foundation Model Embeddings Enable LLM-Based Land Surface Intelligence

Mashrekur Rahman

arXiv:2602.10354v1cs.CLcs.LG

TL;DR

Satellite foundation-model embeddings have limited established physical interpretability, constraining their integration into environmental decision systems. The paper analyzes 64-dimensional AlphaEarth embeddings against 26 environmental variables, tests spatial and temporal robustness, and builds a retrieval-augmented Land Surface Intelligence system. It finds that the embeddings encode structured, physically coherent land-surface information and can be operationalized for environmental and geospatial intelligence.

  • Problem

    Physical interpretability of satellite foundation-model embeddings remains limited, constraining their integration into environmental decision systems.

  • Method

    The study analyzes AlphaEarth embeddings against 26 environmental variables, tests spatial and temporal stability, and applies validated interpretations in a retrieval-augmented system.

  • Results

    AlphaEarth embeddings encode structured, physically coherent land-surface information, with individual dimensions mapping onto specific properties and the embedding space explaining most variance in most variables.

  • Takeaways & Limitations

    The findings support operationalizing satellite foundation-model embeddings for environmental and geospatial intelligence.

  • Takeaways & Limitations

    The analysis is restricted to the 64-dimensional embedding space and 26 environmental variables.

Abstract

from arXiv · show

Satellite foundation models produce dense embeddings whose physical interpretability remains poorly understood, limiting their integration into environmental decision systems. Using 12.1 million samples across the Continental United States (2017--2023), we first present a comprehensive interpretability analysis of Google AlphaEarth's 64-dimensional embeddings against 26 environmental variables spanning climate, vegetation, hydrology, temperature, and terrain. Combining linear, nonlinear, and attention-based methods, we show that individual embedding dimensions map onto specific land surface properties, while the full embedding space reconstructs most environmental variables with high fidelity (12 of 26 variables exceed $R^2 > 0.90$; temperature and elevation approach $R^2 = 0.97$). The strongest dimension-variable relationships converge across all three analytical methods and remain robust under spatial block cross-validation (mean $ΔR^2 = 0.017$) and temporally stable across all seven study years (mean inter-year correlation $r = 0.963$). Building on these validated interpretations, we then developed a Land Surface Intelligence system that implements retrieval-augmented generation over a FAISS-indexed embedding database of 12.1 million vectors, translating natural language environmental queries into satellite-grounded assessments. An LLM-as-Judge evaluation across 360 query--response cycles, using four LLMs in rotating generator, system, and judge roles, achieved weighted scores of $μ= 3.74 \pm 0.77$ (scale 1--5), with grounding ($μ= 3.93$) and coherence ($μ= 4.25$) as the strongest criteria. Our results demonstrate that satellite foundation model embeddings are physically structured representations that can be operationalized for environmental and geospatial intelligence.

1. Introduction

The study addresses limited understanding of whether AlphaEarth embeddings encode physically meaningful environmental features and how those relationships can support environmental information systems. It evaluates interpretability, spatial and temporal robustness, and retrieval-augmented generation for natural-language environmental queries.

  • Research gap: Systematic characterization of AlphaEarth dimension–variable relationships remains limited despite applications in environmental and spatial tasks.The gap concerns whether embedding dimensions correspond to specific land-surface properties and whether the full representation can support broader environmental-system integration.
  • Validation motivation: Spatial autocorrelation can inflate apparent geospatial generalization when random validation allows nearby samples to share information.Spatial block validation is therefore treated as a critical methodological consideration.
  • Application: Validated embedding interpretations are used to develop a Land Surface Intelligence system for natural-language environmental queries through retrieval-augmented generation.The system grounds language-model responses in satellite-derived embedding data.
  • Evaluation: The study also asks how LLM-based geospatial systems should be evaluated for scientific grounding and response quality beyond traditional retrieval metrics.The evaluation target includes response dimensions such as grounding and utility.
  • Study design: The study characterizes a 64-dimensional AlphaEarth space against 26 environmental variables across 12.1 million CONUS samples from 2017–2023.The variables span terrain, climate, vegetation, hydrology, and urban development, with spatial and temporal stability analyses.

2. Methods

The study defines a regular CONUS sampling grid and extracts seven annual AlphaEarth composites, producing approximately 12.1 million samples for analysis.

  • Sampling grid: Approximately 2.34 million CONUS locations were sampled on a regular grid spaced at 0.025° intervals, or about 2.75 km.The grid was bounded by 125.0°W–66.5°W longitude and 24.5°N–49.5°N latitude.
  • Temporal coverage: Seven AlphaEarth annual composites from 2017–2023 yielded approximately 1.73 million samples per year.The annual extraction was performed for every grid location.
  • Dataset assembly: Approximately 12.1 million samples formed the aggregated dataset.This total combines the annual samples across the seven available composites.

2.2. AlphaEarth Embeddings

AlphaEarth embeddings are 64-dimensional annual vectors derived from multimodal satellite observations and extracted at 1 km scale with a 500 m point buffer.

  • Embedding representation: AlphaEarth produces 64-dimensional vectors labeled A00–A63 from a geospatial embedding model trained on multimodal satellite data.The source observations include Sentinel-2 optical imagery, Sentinel-1 radar, Landsat, and ancillary Earth-observation sources.
  • Temporal aggregation: Each annual composite summarizes the full observation time series for a location and year into one embedding vector.This creates a single annual representation for each sampled location.
  • Extraction and preprocessing: Embeddings were extracted at 1 km scale with a 500 m point buffer.The distributed 8-bit quantized values were converted to floats for analysis.
  • Quality control: Samples with missing embedding values caused by edge effects or data gaps were excluded.This exclusion was part of the preprocessing procedure.

2.3. Environmental Variables

The dataset pairs 64 AlphaEarth embedding dimensions with 26 environmental variables spanning terrain, soil, vegetation, temperature, climate, hydrology, and urban conditions across CONUS.

  • Terrain: Terrain variables include elevation, slope, aspect, and log-transformed flow accumulation derived from SRTM terrain data.These terrain variables are temporally static.
  • Soil: Soil variables include clay fraction, organic carbon, pH, and available water capacity from OpenLandMap and SoilGrids.Soil variables were extracted once per location because they are temporally static.
  • Vegetation: Vegetation variables include annual mean and maximum NDVI, mean EVI, mean LAI, and tree cover, primarily from MODIS and Hansen datasets.Albedo was derived from MODIS MCD43A3.
  • Dataset scope: The environmental dataset comprises 26 variables spanning seven thematic categories and is paired with 64 AlphaEarth dimensions across approximately 12.1 million samples.Variables were extracted across CONUS from 2017–2023.
  • Dynamic and anthropogenic variables: Temperature, climate, hydrology, and urban variables combine annual composites with static population density across the study period.The categories include land-surface and air temperature, precipitation, soil moisture, runoff, evapotranspiration, impervious surface, nighttime lights, and population density.

2.4. Interpretability Analysis

The study characterizes relationships between 64 AlphaEarth embedding dimensions and 26 environmental variables using correlation, Random Forest, and Transformer analyses. These complementary methods provide dimension-level interpretability through association and importance patterns.

  • Methods: Three complementary methods characterize embedding–environment relationships: Spearman correlation, Random Forest regression, and a multi-task TabTransformer.The analyses examine linear or monotonic association, nonlinear predictive importance, and Transformer-based importance.
  • Methods: 64 embedding dimensions are evaluated against 26 environmental variables using pooled multi-year data.Spearman coefficients are computed for every dimension–variable pair, and Random Forest models use all 64 dimensions as predictors.
  • Correlation analysis: p < 0.001 for all nonzero correlations at the study’s sample size.For each dimension, the primary corresponding variable is selected from the strongest association.
  • Cross-method interpretation: Concordance compares primary variable assignments across the three methods, with moderate convergence between linear and nonlinear characterizations.A dimension is concordant when at least two methods agree on its primary associated variable.

2.5. Validation

Validation tests whether embedding–variable relationships generalize across spatially separated samples and remain stable over time. The design addresses spatial autocorrelation and interannual variability explicitly.

  • Spatial validation: Random cross-validation may overestimate geospatial performance because nearby samples share information through spatial autocorrelation.The validation procedure therefore applies spatial cross-validation to both Random Forest and Transformer models.
  • Spatial validation: 2° × 2° spatial blocks are assigned to five grouped k-folds so samples from each block remain in one fold.This spatial partitioning is used to assess generalization beyond geographically proximate samples.
  • Temporal validation: Seven annual 64 × 26 correlation matrices are computed from n = 300,000 samples per year across 2017–2023.Each annual profile contains 26 correlation values for every embedding dimension.
  • Temporal validation: Temporal stability is measured as the mean pairwise Pearson correlation between each dimension’s annual correlation profiles.Values near 1.0 indicate consistent dimension–variable relationships over time.

2.6. Dimension Dictionary

The dimension dictionary consolidates outputs from all three interpretability methods into a structured lookup table. It provides the interpretive backbone for translating embedding dimensions into environmental meaning.

  • Dictionary contents: The dictionary records each dimension’s primary Spearman, Random Forest, and Transformer-associated variables.It also stores the corresponding correlation strengths and importance scores.
  • Dictionary contents: The dictionary assigns thematic categories and records two-way and three-way concordance flags.These flags summarize agreement among the interpretability methods.
  • System role: The dictionary serves as the interpretive backbone of the Land Surface Intelligence system.It translates raw embedding values into environmental meaning for downstream system use.

2.7. Land Surface Intelligence System

The Land Surface Intelligence system uses retrieval-augmented generation over AlphaEarth embeddings to answer natural-language environmental queries. It combines location resolution, dimension interpretation, intent classification, similarity search, and structured context for grounded assessments.

  • FAISS vector index: 12.1 million AlphaEarth vectors are indexed for nearest-neighbor retrieval, with sub-millisecond query latency for k-nearest-neighbor searches.Each vector links to coordinates, year, and 26 environmental variables; the index uses FAISS.
  • Similarity retrieval: Embedding-space proximity retrieves physically analogous locations across the Continental United States.The system uses similarity in AlphaEarth space to support comparative environmental analysis without explicit feature matching.
  • Query resolution: Natural-language locations and temporal references are resolved to coordinates and study years before retrieving the corresponding embedding and environmental variables.When no year is specified, the system defaults to the most recent available year.
  • Interpretation and intent: Dimension interpretations link embedding values to environmental meaning, including correlations above |ρ| > 0.5.The interpreted dimensions are combined with one of ten intent categories and relevant environmental variables.
  • Similarity and intent: The system retrieves the k most similar locations, with default k = 10, through the FAISS index.Intent classification determines which variables and interpretations receive emphasis in the response.
  • Retrieval-augmented generation: A structured context document containing retrieved values, dimension interpretations, and similar-location metadata is passed to an LLM through RAG.The LLM synthesizes the satellite-derived information into a natural-language environmental assessment and is instructed to reference retrieved data.

2.8. System Evaluation

The system evaluation uses rotating LLM roles and balanced environmental queries to assess response quality beyond traditional retrieval metrics. Responses are scored with weighted criteria emphasizing grounding, scientific accuracy, completeness, coherence, and practical utility.

  • Evaluation Design: Four LLMs rotated among query generation, system evaluation, and judging, with generator and judge models differing from the system model.This design produced 12 distinct configurations and was intended to prevent self-evaluation bias.
  • Evaluation Design: 360 evaluation cycles combined four system models with three generator/judge configurations, using 30 queries per configuration.The experiment used rotating LLM roles across distinct evaluation configurations.
  • Evaluation Design: Queries sampled CONUS locations uniformly and assigned ten intent types through round-robin allocation for balanced coverage across use cases.Templates introduced natural-language variation, including flood-risk and drought-vulnerability queries.
  • Evaluation Criteria: Five criteria evaluated responses on grounding, scientific accuracy, completeness, coherence, and practical utility using detailed 1–5 scoring rubrics.Grounding checks references to actual embedding data and environmental variables, while practical utility assesses actionability for environmental decisions.
  • Evaluation Criteria: Sweighted = 0.25 · G + 0.25 · A + 0.20 · C + 0.15 · H + 0.15 · U combines the five criterion scores using their assigned weights.G, A, C, H, and U denote grounding, scientific accuracy, completeness, coherence, and practical utility, respectively.
  • Evaluation Procedure: Judge models received the query, location metadata, complete response, and rubrics, then produced criterion scores with reasoning; weighted-score correlations assessed inter-judge reliability.The reliability analysis compared score distributions across the four judge models and overlapping query–response pairs.

3. Results

AlphaEarth embeddings show physically coherent, predictive, spatially generalizable, and temporally stable relationships with environmental variables. These validated representations support a land-surface intelligence system whose responses are generally grounded and coherent, though temporal-change and domain-specific utility remain weaker.

  • Dimension–variable associations: Hierarchical clustering groups temperature, vegetation, and hydrology dimensions, indicating physically coherent organization rather than arbitrary environmental encoding.Spatial maps further show that dimension distributions mirror correlated environmental patterns.
  • Predictive reconstruction: The Transformer exceeds Random Forest for all 26 variables, including annual precipitation rising from R2 = 0.83 to R2 = 0.92.Its attention mechanism also identifies distinct dimension sets supporting vegetation and temperature predictions.
  • Land-surface intelligence: Location profiling and hydrology score highest, whereas temporal-change queries score lowest at μ = 3.56.The weaker temporal performance is consistent with embedding composites lacking explicit multi-year comparison.

4. Conclusion

This study finds that AlphaEarth embeddings encode structured, physically coherent land-surface information and uses validated interpretations to build a Land Surface Intelligence system for environmental queries.

  • AlphaEarth embeddings encode structured, physically coherent information about land-surface properties across climate, vegetation, hydrology, temperature, and terrain.
  • The strongest embedding–environment relationships remain consistent across Spearman correlation, Random Forest regression, and attention-based Transformer analysis.
  • The embedding space reconstructs temperature, elevation, and vegetation indices with particularly high fidelity, while most variables’ relationships generalize across spatial blocks and seven years.
  • Weaker performance and stability identify representation limits for fine-scale terrain properties and urban indicators.
  • Validated interpretations underpin a Land Surface Intelligence system that retrieves environmentally analogous locations through retrieval-augmented generation over a FAISS-indexed embedding database.
  • The system produces satellite-grounded, scientifically coherent responses, linking embedding interpretability to query-level performance.
  • The study is restricted to the Continental United States and seven years of AlphaEarth annual composites, so broader geographic and temporal generality remains to be tested.

Code and Data Availability

The study provides analysis code, supplementary materials, and publicly available input datasets, while the Land Surface Intelligence system code is not publicly released because it depends on authenticated Dartmouth Chat API access.

  • Public materials: Analysis code and supplementary materials are publicly available at the study’s Zenodo repository.The supplementary materials include 30 representative LLM-as-Judge query–response–judgment cycles stratified by intent type, score range, and system model.
  • System-code limitation: The Land Surface Intelligence system code is not included in the public repository because its components depend on the Dartmouth Chat API, which requires institutional authentication credentials.
  • Data availability: All input datasets are publicly available through Google Earth Engine, and the extracted co-located sample dataset is available from the corresponding author upon reasonable request.
Loading 2602.10354v1…