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

TL;DR

This study systematically characterizes Google AlphaEarth satellite foundation model embeddings, showing that the 64‑dimensional space encodes physically meaningful land surface properties across temperature, vegetation, hydrology, and terrain. By employing Spearman correlations, Random Forests, and a multi‑task Transformer, the authors demonstrate strong, robust associations with 26 environmental variables and validate these relationships through spatial block cross‑validation and seven‑year temporal analyses. They then build a Land Surface Intelligence system that uses a FAISS index for retrieval and retrieval‑augmented generation to ground natural language queries in satellite data, with LLMs evaluated via a rotating, multi‑model Judge framework. The results indicate that embedding interpretations are stable and that the system achieves grounding and coherence in responses, suggesting that satellite foundation model embeddings can be operationalized for geospatial intelligence and environmental decision support. The work highlights practical pathways to extend these methods globally, incorporate multi‑year change detection, and improve urban and soil feature representations for broader applicability.

Abstract

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.

Figures (7)

• Figure 1: Land Surface Intelligence system architecture. The pipeline begins with a natural language query that is resolved to geographic coordinates and a target year. The corresponding AlphaEarth embedding vector and environmental variables are retrieved from the FAISS-indexed database of 12.1 million samples. Each embedding dimension is interpreted using the dimension dictionary compiled from the Spearman, Random Forest, and Transformer analyses. The query is classified into one of ten intent categories, and the $k$ most similar locations are retrieved via nearest-neighbor search. The structured context document, containing location data, dimension interpretations, and similar-location metadata, is provided to a large language model through retrieval-augmented generation to produce a grounded environmental assessment.
• Figure 2: AlphaEarth embedding interpretability analysis. (a) Spearman rank correlation matrix ($64 \times 26$) between embedding dimensions and environmental variables, with hierarchical clustering applied to both rows and columns. Color bars along the top axis indicate the thematic category of each environmental variable. (b) The 20 most interpretable dimensions ranked by their strongest absolute Spearman correlation ($|\rho|$), with labels indicating the primary associated variable. Vertical reference lines mark $|\rho| = 0.5$ and $|\rho| = 0.7$. (c) Predictive power ($R^2$, 5-fold cross-validation) of Random Forest and Transformer models for each environmental variable, using the 64-dimensional embedding as input. Variables are sorted by decreasing $R^2$; horizontal reference lines mark $R^2 = 0.5$, 0.7, and 0.9.
• Figure 3: Method convergence analysis. (a) Bipartite network graphs showing the primary dimension--variable connections identified by each of the three interpretability methods (Spearman, Random Forest, Transformer). Nodes on the left represent embedding dimensions and nodes on the right represent environmental variables, colored by thematic category. (b) Method agreement at the pair level, comparing absolute Spearman $|\rho|$ against Random Forest permutation importance across all $64 \times 26 = 1{,}664$ dimension--variable pairs.
• Figure 4: Validation analysis. (a) Spatial generalization: comparison of random and spatial block cross-validation $R^2$ for each environmental variable using the Transformer model, with generalization gaps ($\Delta R^2$) annotated for variables where the gap exceeds 0.02. (b) Temporal stability of all 64 embedding dimensions, measured as mean pairwise Pearson correlation between annual Spearman profiles (2017--2023). Bars are colored by primary thematic category; vertical reference lines mark $r = 0.90$ and $r = 0.95$. (c) Year-to-year profile correlation matrix across the seven study years. (d) Temporal evolution of the top 10 dimension--variable pairs by absolute Spearman $|\rho|$ over 2017--2023.
• Figure 5: Spatial interpretability demonstration. Six selected dimension--variable pairs spanning distinct environmental categories are shown. For each pair, the left panel maps the embedding dimension value and the right panel maps the corresponding environmental variable across CONUS. Spearman $\rho$, Random Forest $R^2$, and Transformer $R^2$ are reported for each pair. The spatial correspondence between embedding dimensions and their associated environmental variables confirms that the learned representations encode geographically coherent physical features.