Persistent Urban Heat Analysis from Multi-Satellite Observations Using Machine Learning

Annual land surface temperature (LST) products are widely used to characterize urban heat, but absolute temperatures combine local surface variation with broader differences among cities and years. This study evaluates AlphaEarth annual satellite embeddings for predicting absolute LST and city-year median-centered thermal anomalies across seven cities from 2017 through 2025. A predictor comparison evaluated the Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), their combination, Landsat surface reflectance, and AlphaEarth embeddings using Random Forest regression. AlphaEarth achieved R2 = 0.812 and RMSE = 2.99 °C for absolute LST prediction and R2 = 0.648 for median-centered anomalies, outperforming conventional predictors. Spatial and temporal validation demonstrated strong recovery of local thermal structure, although transfer to previously unseen cities was weaker. Fixed-location analysis showed strong agreement in multiyear mean anomalies (r = 0.914) but limited recovery of interannual variability (r = 0.096). Geographic reconstruction across seven cities further demonstrated the ability to represent persistent spatial thermal patterns, with greater agreement for mean anomalies than warm-decile recurrence. A supplementary Local Climate Zone (LCZ) analysis associated compact built environments with higher relative surface temperatures and greater hot-tail prevalence, while vegetated environments exhibited cooler conditions. These findings demonstrate the value of AlphaEarth embeddings for characterizing urban surface thermal structure while identifying limitations in temporal variability and geographic transfer.