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

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.

💡 Research Summary

This paper investigates the physical interpretability of Google’s AlphaEarth satellite foundation model embeddings and demonstrates how these embeddings can be operationalized for large‑language‑model (LLM) based land‑surface intelligence. Using a massive dataset of 12.1 million samples covering the contiguous United States from 2017 to 2023, the authors extract 64‑dimensional AlphaEarth embeddings for each 1 km location and pair them with 26 environmental variables spanning terrain (elevation, slope, aspect, flow accumulation), soil (clay fraction, organic carbon, pH, water capacity), vegetation (NDVI, EVI, LAI, tree cover), temperature (day/night land‑surface temperature, air temperature, dew point), climate (annual precipitation, max‑month precipitation), hydrology (soil moisture, runoff, evapotranspiration), and urban metrics (impervious surface, nighttime lights, population density).

Three complementary interpretability techniques are applied:

1. Spearman rank correlation – computes a 64 × 26 matrix of non‑parametric correlations; the variable with the highest absolute ρ for each dimension is taken as its primary associate.
2. Random Forest regression – trains a separate RF model for each environmental variable using all 64 dimensions as predictors; permutation importance identifies the top three dimensions per variable.
3. Multi‑task Transformer – a four‑layer transformer jointly predicts all 26 variables from the embedding vector; gradient‑based input importance and self‑attention weights provide a nonlinear attribution map.

Concordance across methods is defined as at least two methods agreeing on the primary variable for a dimension. Out of 64 dimensions, 48 achieve this two‑method agreement, and 12 achieve full three‑method agreement, indicating a substantial portion of the embedding space is physically grounded. Notable examples include dimension A07 strongly linked to elevation, A22 to mean NDVI, and several dimensions to temperature and precipitation.

Reconstruction performance is evaluated by fitting linear models that map the full 64‑dimensional space to each environmental variable. Twelve variables achieve R² > 0.90, with temperature and elevation reaching R² ≈ 0.97, confirming that the embedding collectively encodes rich physical information.

To guard against spatial over‑fitting, the authors perform spatial block cross‑validation (2° × 2° blocks) with grouped 5‑fold splits. The average drop in R² between random and spatial splits (ΔR²) is only 0.017, demonstrating robust spatial generalization. Temporal stability is assessed by computing yearly Spearman correlation profiles for each dimension; the mean pairwise Pearson correlation across the seven years is 0.963, showing that dimension‑variable relationships are stable over time.

All attribution results are compiled into a “dimension dictionary” that maps each embedding dimension to its most relevant environmental variable(s) along with quantitative scores. This dictionary serves as the knowledge base for a retrieval‑augmented generation (RAG) system.

The RAG system indexes the full set of embeddings with FAISS (IVFFlat, ~3,500 clusters, probe = 64) and stores accompanying metadata (coordinates, year, 26 variables) in Parquet files. When a user poses a natural‑language environmental query, an LLM parses the query, retrieves the k‑nearest embedding vectors, and generates a response grounded in the retrieved satellite‑derived data.

Evaluation uses an “LLM‑as‑Judge” framework: four distinct LLMs rotate through generator, system, and judge roles across 360 query‑response cycles. Responses are scored on grounding, coherence, relevance, and completeness on a 1–5 scale. The system attains an overall mean score of 3.74 ± 0.77, with grounding (3.93) and coherence (4.25) being the strongest dimensions, indicating that the answers are both data‑driven and linguistically fluent.

In summary, the study provides a rigorous, multi‑method validation that AlphaEarth embeddings are physically interpretable and stable across space and time. By leveraging these embeddings within a FAISS‑backed RAG pipeline, the authors deliver a functional land‑surface intelligence platform that can answer natural‑language environmental questions with satellite‑grounded evidence. The work opens pathways for integrating foundation‑model embeddings into decision‑support tools for climate monitoring, disaster response, and resource management.