AlphaEarth Satellite Embeddings for Modelling Climate Sensitive Diseases Towards Global Health Resilience
Usman Nazir, I-Han Cheng, Sara Khalid · 2026
Read AlphaEarth Satellite Embeddings for Modelling Climate Sensitive Diseases Towards Global Health Resilience by Usman Nazir, I-Han Cheng, Sara Khalid (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-10949.txt.
Introduction
Usman Nazir, I-Han Cheng, and Sara Khalid test whether satellite data can help predict child health risks tied to climate. Malaria, childhood lung infection, and poor nutrition together kill over two million children under five each year, mostly where climate swings drive spread, exposure, and food.
Three studies use AlphaEarth satellite embeddings as predictors, covering infection and stunted growth in poor settings where health tracking stays sparse and reactive. Satellite views offer a cheap, wide-area complement to ground data, and the synthesis shows where they truly predict child health. Global-health teams get a grounded read on a promising but poorly tested data source.
Catalog record
| CREATOR | Usman Nazir, I-Han Cheng, Sara Khalid |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18372 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.10949
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AlphaEarth Satellite Embeddings for Modelling Climate Sensitive Diseases Towards Global Health Resilience Usman Nazir, I-Han Cheng, Sara Khalid Abstract Malaria, childhood acute respiratory infection, and child undernutrition together account for over two million deaths annually in children under five, with the burden concentrated in low and middle-income countries where climate variability directly modulates transmission, exposure, and nutritional outcomes ( 3 ; 7 ; 5 ) . Despite this, routine health surveillance in these settings remains sparse and reactive. Satellite-derived representations of the Earth’s surface offer a scalable, low-cost complement to traditional covariates, yet their utility as predictors of population health outcomes is poorly characterised ( 1 ) . We summarise findings from three independent studies that evaluate AlphaEarth Foundations 64-dimensional satellite embeddings as predictors of population health outcomes, with a focus on vulnerable population groups e.g. children and populations in low-resource settings. The studies span two health domains: infectious disease (malaria, respiratory infection) and stunting. 1 1 1 Following common usage in the child-nutrition community, we use “stunting” as the umbrella term for chronic child undernutrition. The specific outcome modelled in Case 3 is the Weight-for-Height Z-score (WHZ), which is the continuous variable from which the binary wasting indicator (WHZ < < − 2 -2 ) is derived. Stunting (low height-for-age) and wasting (low weight-for-height) are related but distinct dimensions of undernutrition; our pipeline is agnostic to which of the two is selected as the target. In each study, embeddings provide meaningful predictive value when merged at sufficient spatial granularity: (i) in malaria case prediction across Nigeria, AlphaEarth Embeddings yields consistent per-region R 2 R^{2} gains; (ii) in childhood acute respiratory infection (ARI) prediction across 11 DHS countries, embeddings increase pooled R 2 R^{2} from 0.157 to 0.206, an effect that is robust across three tree-based estimators; (iii) in stunting prediction across 35 countries, a country-level prototype is—as expected—neutral, because country-broadcast embeddings are collinear with country fixed effects. The stunting case is currently blocked on DHS cluster-level coordinate access; we view this as the next and most informative experiment. We close with a request: direct access to the Google Earth AI foundation-model suite, including Population Dynamics embeddings that incorporate health indicators, would substantially accelerate this line of work across all three outcomes. (a) Per-region 2024 test R^{2} across Nigerian states. Each spoke is one state; the inner polygon is the climate-only baseline and the outer polygon is the same model with the 64-dim AlphaEarth fingerprint appended. The outer polygon strictly dominates on every spoke, indicating that the lift is geographically uniform rather than driven by a few high-burden states. (b) Top-20 Random Forest feature importances. Bars are coloured by feature group (baseline climate / contextual vs. AlphaEarth band A_{i} ). Figure 1 : Case 1 — Malaria case prediction in Nigeria (NMEP, 2000–2024; train 2000–2023, test 2024). AlphaEarth embeddings provide a geographically uniform R^{2} gain (left) and emerge as the dominant feature group in the importance decomposition (right), supporting the interpretation that static landscape structure carries malaria-transmission signal not captured by monthly climate covariates. 1 Motivation Malaria, acute respiratory infection, and undernutrition are among the leading causes of child mortality globally, collectively responsible for a disproportionate share of the disease burden in sub-Saharan Africa and South Asia ( 3 ) . Climate variability is a well-established driver of all three: temperature and rainfall govern mosquito breeding and Plasmodium falciparum transmission ( 4 ; 7 ) , ambient air quality and indoor pollution modulate respiratory infection risk, and seasonal drought and flood cycles affect food security and nutritional status ( 5 ) . As climate change intensifies these exposures, the need for scalable, near-real-time environmental monitoring tools for health is increasingly urgent ( 6 ) . Child and population health in LMICs is driven by environmental and structural factors, such as climate, vegetation, infrastructure, land cover, water which are difficult to capture with traditional survey covariates. Geolocated Demographic and Health Survey (DHS) data provides cluster-level outcomes at 2–5 km spatial precision, creating a natural opportunity to test whether learned representations of the Earth’s surface carry predictive signal for health outcomes. We evaluate AlphaEarth Foundations V1 (64-dim, 10-m annual unit-norm embeddings via Google Earth Engine) on three distinct LMIC outcomes spanning two health domains: • Malaria (vector-borne infectious disease): Nigeria, 2000–2024. • Childhood ARI (respiratory infectious disease): 11 DHS countries, 2017–2022. • Child undernutrition (WHZ) (chronic / stunting): 35 DHS countries, 2015+. Each study uses an appropriate baseline (climate-only, pollution-only, or regression-based tabular respectively), and tests whether adding AlphaEarth embeddings improves test-set prediction. The three case studies are presented sequentially below. 2 Case 1 — Malaria Prediction, Nigeria Setup. Annual malaria case counts from Nigeria’s National Malaria Elimination Programme (NMEP) surveillance across 28,209 locations, 2000–2024, linked to monthly ERA5-Land climate (temperature, rainfall, NDVI), contextual covariates, and a static AlphaEarth fingerprint per location. Training is performed on 2000–2023 and held-out evaluation on 2024, giving a one-year forward gap that stresses the model’s ability to generalise beyond the training window. Plasmodium falciparum prevalence ( pfir ) is explicitly excluded from inputs to avoid circular prediction. Result. On the 2024 test year, adding AlphaEarth raises test R^{2} from 0.201 to 0.245. All Nigerian regions show positive \Delta R^{2} in Figure 1(a) (Also see Figure 4(a) in appendix A ). 3 Case 2 — Childhood ARI, 11 DHS Countries Setup. 9,271 DHS clusters across India, Nigeria, Pakistan, Kenya, Nepal, Mozambique, Bangladesh, Cambodia, Lebanon, Tajikistan, and the Philippines (2017–2022). Target: observed number of ARI cases among children under five per cluster. Baseline: gaseous pollutants (CO, NO 2 , SO 2 ) plus controls (elevation, population, temperature, urban/rural). Feature augmentation: 64-dim AlphaEarth embedding sampled at each cluster’s survey year over a 2–5 km buffer. Evaluation: 5-fold \times 2-repeat cross-validation, tested with three tree-based estimators. Result. At the pooled level (Figure 5 in appendix), all three estimators show a consistent monotonic improvement: baseline R^{2}\approx 0.15 –0.17 \rightarrow AlphaEarth-only \approx 0.18 –0.19 \rightarrow both features combined \approx 0.21 . The cross-model consistency of the effect (best: XGBoost, R^{2}=0.210 ) indicates the signal is in the embeddings, not an artefact of any particular inductive bias. Per-country (Figure 2(a) ), embeddings help in all six countries with n\geq 139 clusters (8,696 of 9,271 clusters, 94% of the data), with largest absolute gains in Mozambique ( \Delta R^{2}=+0.096 ) and Nigeria ( +0.093 ). In the smallest- n countries ( n<130 ) the 64-dim embedding causes overfitting under HistGradientBoosting; Random Forest substantially attenuates this and is recommended in data-scarce settings. (a) Per-country performance for the six DHS countries with n\geq 139 clusters (India, Pakistan, Nigeria, Kenya, Nepal, Mozambique; 8,696 of 9,271 clusters total). Left panel: RMSE relative to baseline (lower is better). Right panel: test R^{2} (higher is better). The three polygons compare the gas+controls baseline (red), AlphaEarth alone (orange), and the combined model (green). The combined model strictly dominates baseline on every spoke; the largest absolute gains are in Mozambique ( \Delta R^{2}=+0.096 ) and Nigeria ( +0.093 ). (b) Top-20 permutation importances on a 25% held-out split (combined embedding + gaseous pollutants + controls model). Bars coloured by feature group: AlphaEarth (green), gaseous pollutants (red), controls (blue). Elevation leads, but seven of the top ten features are AlphaEarth bands, and two