📊 statistics

AlphaEarth Satellite Embeddings for Modelling Climate Sensitive Diseases Towards Global Health Resilience

This paper evaluates the utility of AlphaEarth's 64-dimensional satellite embeddings as scalable predictors for climate-sensitive diseases and child undernutrition in low- and middle-income countries, demonstrating significant improvements in predicting malaria and respiratory infections while highlighting data limitations that currently constrain stunting modeling at the cluster level.

Original authors: Usman Nazir, I-Han Cheng, Sara Khalid

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Usman Nazir, I-Han Cheng, Sara Khalid

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 the Earth as a giant, complex puzzle. For a long time, scientists trying to predict where children might get sick (like malaria, bad coughs, or malnutrition) have only been looking at a few specific puzzle pieces: the weather, the temperature, and the rain. They've been trying to guess the picture of public health using just these "weather" pieces.

This paper is about a team of researchers at Oxford who decided to try a different approach. They used a super-smart AI tool called AlphaEarth. Think of AlphaEarth not just as a weather report, but as a high-definition "fingerprint" of the landscape. It looks at the ground from space and captures the static structure of the world—the shape of the land, the type of vegetation, the layout of the soil, and the general "vibe" of a place that doesn't change day-to-day.

The researchers asked: "If we give the AI this full landscape fingerprint, can it predict child health problems better than just looking at the weather?"

They tested this idea in three different scenarios, like running three different experiments in a science lab:

Experiment 1: The Malaria Map (Nigeria)

The Setup: They looked at malaria cases across Nigeria over 24 years.
The Analogy: Imagine trying to predict where mosquitoes will breed. You know rain helps them, but you also know that the shape of the land (like a hidden valley or a specific type of forest) matters, even if it's not raining right now.
The Result: When they added the AlphaEarth "landscape fingerprint" to their model, their predictions got significantly better. It wasn't just a little bit better; it was a clear improvement across the entire country. The AI realized that the static structure of the land holds secret clues about malaria that the monthly weather data misses.

Experiment 2: The Coughing Children (11 Countries)

The Setup: They looked at acute respiratory infections (bad coughs and lung infections) in children across 11 different countries, including India, Nigeria, and Kenya.
The Analogy: Think of air pollution as the "smoke" in a room. Usually, we just measure the smoke. But this team asked, "Does the architecture of the room (the walls, the location, the surrounding environment) also tell us if a child is likely to get sick?"
The Result: Again, adding the landscape fingerprint helped. The model's accuracy jumped up. The AI found that the "fingerprint" of the location was just as important as the pollution levels. It worked consistently across different types of math models, proving the signal was real and not just a fluke.

Experiment 3: The Malnutrition Puzzle (35 Countries)

The Setup: They tried to predict child undernutrition (specifically "stunting," or being too short for one's age) across 35 countries.
The Analogy: This is where they hit a wall. Imagine trying to describe a specific house by only knowing the country it is in. If you tell the AI, "This house is in France," and the AI already knows "France," adding the fact that "This house is in France" doesn't give any new information. It's redundant.
The Result: When they tried to use the landscape fingerprint for the entire country at once, it didn't help at all. The AI said, "I already know the country's general vibe; this new data is just repeating what I know."
The Lesson: The researchers realized they were looking at the wrong scale. To predict malnutrition, they need to look at the specific neighborhood (the cluster), not the whole country. They are currently waiting for permission to access the specific GPS coordinates of the survey locations to try this again at the neighborhood level.

The Big Takeaway

The main lesson from this paper is about zooming in.

When you zoom in to the specific neighborhood or village level (like in the Malaria and Coughing studies), the "landscape fingerprint" is incredibly useful. It acts like a secret decoder ring that reveals health risks hidden in the terrain.
When you zoom out too far to the whole country level (like in the first Malnutrition attempt), the fingerprint becomes useless because it just repeats what we already know about the country.

What they are asking for:
The team is asking Google (who made the AlphaEarth tool) for two things to help them finish their work:

Direct access to a specific part of the AI that tracks population and health data, so they can combine it with the landscape data.
Permission to use the specific GPS coordinates of the survey clusters so they can run the "neighborhood-level" experiment for malnutrition, which they believe will be the final piece of the puzzle.

In short: Satellite "fingerprints" of the land are a powerful new tool for predicting child health, but only if you use them to look at specific neighborhoods, not just entire countries.

Technical Summary: AlphaEarth Satellite Embeddings for Modelling Climate Sensitive Diseases

1. Problem Statement

Malaria, childhood acute respiratory infections (ARI), and child undernutrition collectively account for over two million annual deaths in children under five, with the burden concentrated in low- and middle-income countries (LMICs). Climate variability is a primary driver of these outcomes, yet routine health surveillance in these regions remains sparse and reactive. Traditional health models often rely on survey covariates that struggle to capture complex environmental and structural factors (e.g., vegetation, land cover, water infrastructure). While satellite-derived data offers a scalable alternative, its utility as a predictor for population health outcomes remains poorly characterized.

The core problem addressed is whether learned, high-dimensional representations of the Earth's surface (specifically AlphaEarth Foundations embeddings) can serve as effective predictors for health outcomes in LMICs, and at what spatial granularity these embeddings provide meaningful signal.

2. Methodology

The authors evaluate AlphaEarth Foundations V1, a set of 64-dimensional, 10-meter annual unit-norm embeddings generated via Google Earth Engine. The study spans three distinct health outcomes across two domains: infectious disease and undernutrition.

Experimental Design

Three independent case studies were conducted, each employing a baseline model and testing the incremental value of adding AlphaEarth embeddings:

Case 1: Malaria Prediction (Nigeria)
- Data: 28,209 locations (2000–2024) from the Nigeria National Malaria Elimination Programme (NMEP).
- Baseline: Monthly ERA5-Land climate variables (temperature, rainfall, NDVI) and contextual covariates.
- Augmentation: A static 64-dim AlphaEarth fingerprint per location.
- Validation: Training on 2000–2023; held-out test on 2024 to assess generalization beyond the training window. Plasmodium falciparum prevalence was explicitly excluded to prevent circular prediction.
Case 2: Childhood ARI Prediction (11 LMICs)
- Data: 9,271 Demographic and Health Survey (DHS) clusters across 11 countries (2017–2022).
- Target: Observed number of ARI cases in children under five per cluster.
- Baseline: Gaseous pollutants (CO, NO2, SO2) plus controls (elevation, population, temperature, urban/rural status).
- Augmentation: 64-dim AlphaEarth embeddings sampled at each cluster's survey year over a 2–5 km buffer.
- Validation: 5-fold × 2-repeat cross-validation tested across three tree-based estimators (Random Forest, HistGradientBoosting, XGBoost).
Case 3: Stunting/Undernutrition Prediction (35 LMICs)
- Data: 389,035 children under five across 35 DHS countries (2015+).
- Target: Weight-for-Height Z-score (WHZ), used as a proxy for chronic undernutrition.
- Baseline: Tabular covariates (temperature, age, altitude, income, sanitation, etc.) and an 8-dim country embedding.
- Prototype Test: A country-level AlphaEarth fingerprint was created by averaging 30 population-weighted points per country and broadcasting this single vector to all children within that country.
- Note: A full cluster-level extraction (using ~45,775 unique DHS clusters) was proposed but blocked due to lack of access to DHS cluster coordinates.

3. Key Results

Case 1: Malaria (Nigeria)

Performance: Adding AlphaEarth embeddings increased the test $R^2$ from 0.201 to 0.245 ( $\Delta R^2 = +0.044$ ).
Spatial Distribution: The improvement was geographically uniform across all 28,209 locations; no single high-burden state drove the gain.
Feature Importance: In Random Forest models, AlphaEarth bands emerged as the dominant feature group, surpassing monthly climate covariates. This suggests static landscape structure carries transmission signals not captured by dynamic climate data.

Case 2: Childhood ARI (11 Countries)

Performance: At the pooled level, embeddings increased the $R^2$ from 0.157 to 0.206 ( $\Delta R^2 = +0.049$ ).
Robustness: The improvement was consistent across all three tree-based estimators, indicating the signal resides in the embeddings rather than a specific model's inductive bias.
Granularity: Embeddings improved performance in all six countries with sufficient sample sizes ( $n \ge 139$ clusters). The largest absolute gains occurred in Mozambique ( $\Delta R^2 = +0.096$ ) and Nigeria ( $\Delta R^2 = +0.093$ ).
Feature Importance: In the combined model, seven of the top ten features were AlphaEarth bands. Two specific bands (A50, A60) out-ranked the strongest pollutant feature (NO2).

Case 3: Stunting (WHZ)

Prototype Result (Country-Level): The country-broadcast AlphaEarth fingerprint yielded no meaningful gain ( $\Delta R^2 \approx 0$ ). The maximum observed gain was +0.020 in Lesotho, with no country exceeding a substantive threshold of +0.01.
Explanation: This negative result was expected. A feature constant within a country is collinear with country fixed effects (or the learned country embedding) already present in the model. The embedding absorbed information already carried by country identity rather than adding new signal.
Implication: This confirms that for undernutrition, cluster-level extraction (capturing within-country variation like urban/rural or coastal/inland differences) is structurally necessary for the embeddings to be informative.

4. Significance and Claims

The paper claims that AlphaEarth Foundations embeddings constitute a promising and generalizable feature class for population health modeling in low-resource settings, provided they are applied at the appropriate spatial granularity.

Spatial Granularity is Critical: The utility of the embeddings is strictly dependent on the merge grain. They add value when they can express within-country variation (cluster or location level). When forced to the country level, they become redundant with country fixed effects.
Static vs. Dynamic Signals: The results suggest that static landscape structures (captured by annual embeddings) encode transmission and exposure signals for malaria and ARI that are not captured by conventional monthly climate or pollution covariates.
Scalability: The approach offers a scalable, low-cost complement to traditional covariates, potentially enabling better disease surveillance, early warning systems, and targeted public health interventions.

5. Limitations and Future Work

The authors explicitly note several limitations:

Spatial Displacement: DHS cluster coordinates are displaced (up to 5km urban, 10km rural) for privacy, introducing measurement error that likely attenuates observed effects.
Temporal Resolution: AlphaEarth embeddings are static annual composites and cannot capture intra-annual dynamics (e.g., outbreak-scale variation) without monthly or seasonal versions.
Metric Specificity: The stunting analysis used Weight-for-Height Z-score (WHZ) as a proxy; performance on Height-for-Age (stunting proper) remains unevaluated.
Heterogeneity: Country-level pooling may mask heterogeneity across agroecological zones.

Future Directions:
The authors request direct access to the Google Earth AI foundation-model suite, specifically:

Population Dynamics Embeddings: To complement landscape data with health and socio-demographic indicators.
Cluster-Level Extraction: To enable the definitive test of embeddings on undernutrition (Case 3) by accessing ~45,775 cluster-centered fingerprints.
Methodological Feedback: On weighting strategies (population density vs. uniform) and temporal composites (multi-year means vs. survey-year point estimates).