Daily pan evaporation reconstruction for 1300+ stations in China from 2000-2024

Here we developed a reproducible framework that links data homogenization, daily reconstruction, and process attribution to revisit the evaporation paradox across China’s climate gradients during 2000–2024. Historical D20 observations were first converted to E601-equivalent values using a meteorology-dependent random forest model, yielding a homogenized daily dataset for 1,351 stations. We then reconstructed daily Epan using a Transformer driven only by meteorological forcing. The Transformer outperformed random forest, XGBoost, and LightGBM, achieving the best daily skill (NSE = 0.76).