💬 NLP

Characterizing AlphaEarth Embedding Geometry for Agentic Environmental Reasoning

This paper characterizes the non-Euclidean manifold geometry of Google AlphaEarth's land surface embeddings, revealing significant local curvature and concept direction rotation, and leverages these insights to build an agentic system that outperforms parametric-only reasoning by utilizing embedding retrieval for physically coherent environmental analysis.

Original authors: Mashrekur Rahman, Samuel J. Barrett, Christina Last

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mashrekur Rahman, Samuel J. Barrett, Christina Last

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, magical library that contains a "fingerprint" for every single square inch of the United States. These fingerprints aren't made of ink; they are 64-digit codes generated by a super-smart AI called AlphaEarth. These codes describe everything about a place: how wet the soil is, how hot it gets, what kind of plants grow there, and even the shape of the mountains.

For a long time, scientists knew these codes existed and could use them to answer simple questions like, "Is it raining in Kansas?" But they didn't really understand the shape of the library itself. They didn't know if the codes were arranged in a straight, flat hallway or a twisting, curving maze.

This paper is like a team of cartographers who decided to map that maze. They asked: If we try to do math with these codes (like adding "wetness" to a dry place), will it work? And can we build a robot assistant that uses this map to answer complex environmental questions?

Here is the breakdown of their discovery using some everyday analogies:

1. The Map is Curved, Not Flat

In the world of text (like how computers understand words), we often assume the "space" is flat. If you take the word "King," subtract "Man," and add "Woman," you get "Queen." It works because the map is a straight line.

The researchers found that the AlphaEarth map is not flat. It's more like a crumpled piece of paper or a mountainous terrain.

The Analogy: Imagine trying to draw a straight line on a globe. If you walk in a "straight" line from New York to London, you aren't actually walking in a straight line on the map; you're curving over the Earth.
The Finding: The "direction" of "wetness" in a desert is totally different from the "direction" of "wetness" in a swamp. If you try to do math to make a dry place wet by just adding a "wetness vector," the math fails because the map twists and turns. The rules change depending on where you are standing.

2. The "Local" vs. "Global" Confusion

The researchers discovered that while there are some big, general trends (like a "temperature axis" for the whole country), these trends don't work well for specific neighborhoods.

The Analogy: Think of a weather forecast. A "Global" forecast might say, "It's summer in the US." But if you are in a specific valley in Colorado, the local weather is dictated by the mountain peaks right next to you, not the general summer trend.
The Finding: The AI's understanding of a place is highly local. A "temperature" code in the mountains points in a different direction than a "temperature" code in the plains. This means you can't just use one universal rulebook for the whole country.

3. The Robot Assistant (The Agentic System)

Since doing math on these codes is risky (because the map is so curvy), the researchers built a new kind of robot assistant. Instead of trying to invent new answers by doing math, this robot acts like a librarian.

How it works: When you ask, "Compare the flood risk in Portland and Phoenix," the robot doesn't guess. It goes to the library, finds the exact "fingerprint" for Portland, finds the one for Phoenix, and compares the real data side-by-side.
The Secret Weapon: The robot also has a "confidence meter." Because the map is curvy, the robot knows that in some places (like flat plains), the library is very reliable. In other places (like jagged mountains), the library is a bit messy. The robot tells you, "I'm very sure about this answer," or "This answer is a bit fuzzy because the terrain is complex."

4. The "Smartness" of the Reader Matters

Here is the most surprising part: The value of this detailed map depends on how smart the robot reading it is.

The Analogy: Imagine giving a very detailed, complex map to a toddler versus a seasoned explorer.
- The Toddler (Less Advanced AI): If you give the toddler the complex map, they get confused. They trip over the details and give a worse answer than if you just gave them a simple list.
- The Explorer (More Advanced AI): If you give the seasoned explorer the same complex map, they use it perfectly. They navigate the twists and turns to give a much better, more accurate answer.
The Finding: The researchers tested two different AI models. The smarter one (Opus) used the complex map to get better answers. The slightly less smart one (Sonnet) got slightly confused by the extra details and performed worse.

The Big Takeaway

This paper teaches us that Earth observation data is messy and curved, not neat and straight.

To build AI that can reason about our planet, we shouldn't try to force the data into simple math equations. Instead, we should build retrieval systems (like a librarian) that look up real data. Furthermore, we need to give these systems "confidence meters" so they know when the data is reliable. Finally, the more complex the map we give them, the smarter the AI needs to be to use it effectively.

In short: Don't try to do math on a mountain; just hire a guide who knows the terrain.

1. Problem Statement

Earth observation foundation models (e.g., Google AlphaEarth) compress multispectral satellite imagery into dense embedding vectors (e.g., 64 dimensions) to represent land surface information. While these models perform well on specific tasks like land cover classification, the geometric structure of their embedding spaces remains poorly understood.

Key gaps addressed by this paper include:

Geometric Nature: Is the embedding space flat (Euclidean) or curved (non-Euclidean)? How do local variations relate to global structures?
Operational Viability: Can standard vector arithmetic operations (analogous to NLP word embeddings, e.g., $king - man + woman = queen$) be applied to environmental embeddings to perform compositional reasoning?
System Design: How can this geometric understanding inform the design of agentic systems that reason over satellite data?
Model Dependency: Does the utility of geometric metadata depend on the reasoning capabilities of the Large Language Model (LLM) consuming the data?

2. Methodology

The study utilizes 12.1 million 64-dimensional embedding vectors from Google AlphaEarth covering the Continental United States (CONUS) from 2017–2023, co-located with 26 environmental variables (e.g., temperature, precipitation, soil). The research proceeds in three phases:

Phase 1: Manifold Characterization

Global Structure: Eigendecomposition of the covariance matrix was performed to calculate the Participation Ratio (PR), measuring effective dimensionality.
Intrinsic Dimensionality: The Levina–Bickel Maximum Likelihood Estimator (MLE) was used to estimate local intrinsic dimensionality at 200,000 locations.
Local Geometry: Local Principal Component Analysis (PCA) was conducted at 10,000 probe locations to measure:
- Alignment between local and global principal components ( $|\cos \theta|$ ).
- Tangent space rotation (instability) between adjacent probes.
- Multi-scale behavior by varying neighborhood sizes ( $k=20$ to $2000$).

Phase 2: Testing Compositional Operations

Vector Arithmetic: Three experiments tested if embeddings could be manipulated algebraically:
1. Targeted Shift: Moving an embedding along a direction to increase a specific property (e.g., making a location "wetter").
2. Property Transfer: Transplanting a property from one location to another while preserving others.
3. Analogy: Testing geospatial analogies (e.g., $A - B + C \approx D$ ).
Direction Identification: Directions were identified via unsupervised PCA and supervised Ridge Regression (linear probes) at global, regional, and local scales.
Retrieval Coherence: Measured the physical consistency of $k$ -nearest neighbors in embedding space and built a confidence model to predict retrieval reliability based on geometric features.

Phase 3: Agentic System Development & Evaluation

Architecture: A ReAct-style agentic system was built using Claude Sonnet 4.5. It integrates nine tools:
- 5 Retrieval Tools: Location resolution, embedding retrieval, similarity search, dimension interpretation, and location comparison.
- 4 Geometry-Aware Tools: Retrieval confidence assessment, regional profile lookup, geometric context (dimensionality/curvature), and similar region identification.
Evaluation: 120 queries across three complexity tiers (Single-location, Multi-step comparison, Counterfactual) were evaluated using an LLM-as-Judge framework (Gemma-3-27B) on five criteria (Grounding, Accuracy, Completeness, Coherence, Utility).
Ablation & Benchmarking: Compared the full system against baselines (No geometric tools, Deterministic pipeline, LLM-only) and tested cross-model performance using Claude Opus 4.6.

3. Key Contributions

Geometric Characterization of AlphaEarth: First comprehensive mapping of the embedding manifold, revealing it is non-Euclidean, curved, and locally heterogeneous.
Refutation of Vector Arithmetic: Demonstrated that compositional vector arithmetic fails in this space because concept directions rotate significantly across the manifold, unlike in NLP.
Geometry-Aware Agentic System: Developed a novel agent that uses geometric metadata (curvature, dimensionality, confidence) to guide retrieval and reasoning, rather than relying solely on parametric knowledge.
Model-Capability Scaling: Provided evidence that the value of geometric metadata is not static; it scales with the reasoning capability of the underlying LLM.

4. Key Results

A. Manifold Geometry

Effective Dimensionality: The 64-dimensional space has an effective dimensionality of 13.3 (Participation Ratio) globally and a local intrinsic dimensionality of ~10.4.
Curvature: The manifold is highly curved. 84% of sampled locations have tangent space angles exceeding 60° between neighbors.
Misalignment: The alignment between local and global principal components is weak ( $|\cos \theta| = 0.17$ ), barely above the random baseline (0.125). This implies a single global dictionary cannot describe local variations.
Stability: The geometric structure is temporally stable across the 7-year period.

B. Compositional Operations vs. Retrieval

Arithmetic Failure: Vector arithmetic (shifting, transfer, analogy) yielded poor precision. Even supervised linear probes, which achieved high predictive accuracy ( $R^2 > 0.9$ ), identified directions that rotated across the manifold, failing to generalize.
Retrieval Success: In contrast, retrieval produced physically coherent results. A linear model could predict retrieval coherence ( $R^2 = 0.32$ ) using geometric features (e.g., distance to neighbors, intrinsic dimensionality), allowing the agent to estimate confidence without ground truth.

C. Agentic System Performance

Retrieval Dominance: Satellite embedding retrieval is the primary driver of quality. The full system ( $\mu = 3.79$ ) significantly outperformed the LLM-only baseline ( $\mu = 3.03$ ).
Multi-Step Reasoning: The agentic architecture excelled at Tier 2 (Multi-step comparison) queries ( $\mu = 4.28$ ), a task class that deterministic pipelines could not handle.
Geometric Tools: For the weaker model (Sonnet 4.5), adding geometric tools slightly reduced the weighted score (due to planning complexity) but increased Geometric Grounding scores.

D. Cross-Model Benchmark (Sonnet vs. Opus)

Model Dependency: The utility of geometric tools depends on the model's reasoning capability.
- Sonnet 4.5: Geometric context reduced performance ( $\Delta = -0.12$ ) due to planning overhead.
- Opus 4.6: Geometric context improved performance ( $\Delta = +0.07$ ) and significantly boosted Geometric Grounding scores ($3.38$ vs. $2.64$).
Conclusion: Stronger models can effectively utilize complex geometric metadata to refine reasoning, while weaker models struggle with the added planning burden.

5. Significance

This paper fundamentally shifts the approach to using satellite foundation models from algebraic manipulation (treating embeddings like NLP vectors) to retrieval-based reasoning informed by geometric structure.

Theoretical Impact: It establishes that environmental embedding spaces are curved manifolds where local geometry dictates interpretation, invalidating the assumption of global linearity.
Practical Impact: It provides a blueprint for building Agentic Geospatial Intelligence systems. By incorporating geometric metadata (confidence, dimensionality, regional profiles), these systems can make more reliable, context-aware environmental assessments.
Future Direction: The findings suggest that as LLMs become more capable, the value of detailed geometric characterization of embedding spaces will increase, enabling more sophisticated, multi-step environmental reasoning that deterministic pipelines cannot achieve.