Transformer-based mapping of 10 m spatiotemporally seamless daily urban evapotranspiration using geospatial foundation embeddings

Urban evapotranspiration (ET) plays a critical role in regulating urban heat and governing water–energy exchanges. However, high-resolution ET estimation across entire cities remains challenging due to pronounced surface heterogeneity, which limits the applicability of physical models originally developed for homogeneous natural surfaces. Moreover, satellite-derived ET estimates are often spatially and temporally fragmented due to their dependence on clear-sky conditions and long revisit intervals. To address these limitations, we propose a transformer-based framework for generating spatiotemporally seamless, daily urban ET maps at 10 m resolution. The framework integrates 10 m vegetation greenness from Sentinel-2 NDVI, Google satellite embedding data, and gap-free daily meteorological variables (precipitation, air temperature, and radiation etc.). A key advantage of the proposed method is its use of a group-to-token transformer encoder, which learns nonlinear interactions among atmospheric demand, surface structure, and vegetation dynamics without requiring explicit urban parameterization. The model is trained on time-series observations from 64 globally distributed FluxNet towers and an urban flux site in Shenzhen, where the flux data are gap-free and suitable for continuous daily modeling. The trained model performs well, with an overall R² of 0.92 and RMSE of 0.37 mm d⁻¹. Comparative analyses further indicate that the proposed framework reduces RMSE by approximately 16% and improves spatial discrimination in heterogeneous urban settings. The trained model was applied to Shenzhen, China, generating a spatiotemporally seamless daily urban ET dataset at 10 m resolution for 2017–2024. Validation confirms the reliability of the approach for urban ET mapping, with a mean R² of 0.56 and RMSE of 0.45 mm d⁻¹. Overall, this study establishes a practical pathway for producing cloud-gap-free, 10 m daily urban ET estimates by integrating routinely available reanalysis data with satellite-derived geospatial embeddings. The framework robustly captures intra-urban heterogeneity, supports seasonal to interannual analyses and pixel-wise trend detection, and enables scalable applications in urban hydroclimate research and planning.