AlphaEarth Embeddings Improve Malaria Forecasting in Low-Resource Settings

Malaria remains a leading cause of preventable child death in low- and middle-income countries, and its transmission is strongly climate-sensitive. Yet routine surveillance is weakest exactly where the burden is heaviest, leaving the settings that most need forecasts with the least data to build them. Geospatial foundation models, trained on global Earth-observation data, offer a promising alternative, encoding landscape structure that standard climate covariates do not capture. In this study, we evaluate whether embeddings from one such model, AlphaEarth Foundations, improve malaria case prediction beyond a climate-covariate baseline. Using annual case counts from the Malaria Atlas Project in Nigeria and India under a one-year forward holdout, we find that appending the embedding raises the coefficient of determination (test $R^2$) from $0.623$ to $0.777$ in Nigeria and from $0.867$ to $0.881$ in India. The approach is a scalable input for malaria early warning precisely where climate change is redrawing the transmission map: it runs on already-public satellite layers and needs no new ground data.