Downscaling urban land surface temperature using AlphaEarth satellite GeoAI embeddings: A cross-city evaluation in the US

Fine-scale land surface temperature (LST) data are critical for urban heat management, yet open-source LST datasets have long been limited to 30–500 m resolution. Recently, Google released its AlphaEarth Foundations (AEF), a 64-dimensional GeoAI embedding dataset covering the entire globe at a spatial resolution of 10 m. This study introduces a novel application and comprehensive evaluation of the AEF embedding dataset for downscaling 30-m Landsat LST to 10-m resolution in urban environments using deep learning. As case studies, LST downscaling was conducted across 26 US cities using an artificial neural network that takes the embeddings as input to predict LST values and is trained with spatially stratified samples. We achieved a mean root-mean-squared error (RMSE) of 1.57 °C and a mean R 2 of 0.733 on the test set. Spatial cross-validation reveals minimal deviations (standard deviation of RMSE ∼0.0012 on the log-transformed scale) across the folds. The prediction errors were not spatially clustered and had no significant hotspots for any of the cities. Further, SHAP analysis demonstrated that AEF layers 31, 18, and 1 had the highest feature importance for predicting LST across the cities. Finally, the model trained for each city was applied to predict LST in other cities to assess the model's cross-city geographic transferability. Model transferability succeeds across places with similar forest structure or aridity but is hindered by extreme temperatures and the distinct characteristics of tropical ecosystems. The downscaled high-resolution LST product can be used to monitor urban thermal stress and inform neighborhood-level climate adaptation planning.