High Spatiotemporal Mapping of Urban Evapotranspiration via a Hybrid Physics-Based and Machine Learning Framework

High-resolution mapping of urban evapotranspiration (ET) is challenged by landscape heterogeneity and complex natural-anthropogenic interactions, with a critical gap in fine-scale ET products that capture these land-atmosphere dynamics. This study bridges this gap by developing a hybrid physics-based and machine learning framework to generate 10-meter, hourly urban ET maps for Wuhan, China. We first simulate hourly latent heat flux at 333-m resolution using the Weather Research and Forecasting (WRF) model to establish a physical background field. Concurrently, high-resolution (10-m) surface features, including the Normalized Difference Vegetation Index (NDVI) and urban morphological parameters, are derived from Sentinel-2 imagery. Spatial downscaling from 333 m to 10 m is achieved by leveraging eddy covariance data from Wuhan’s urban flux tower. Using flux footprint modeling, hourly tower measurements are linked to the fine-scale land-cover configuration of their source areas, establishing a physical relationship between latent heat flux and surface properties. A Random Forest model is trained on this relationship and applied citywide to generate 10-m hourly ET maps. The resulting dataset effectively resolves intra-urban variability, clearly capturing sharp ET gradients between impervious surfaces and green spaces. This work provides a scalable and physics-grounded pathway for high-fidelity urban ET mapping, offering valuable insights for urban heat mitigation and water resources management.