A high-resolution (1d, 9km) and long-term (1950-2022) gridded evapotranspiration dataset

Evapotranspiration (ET) is the second largest hydrological flux over land surface and connects water, energy, and carbon cycles. Quantifying spatio-temporal ET variability remains greatly challenging due to limited site observations and significant model uncertainties. To address this issue, we develop a multimodal machine-learning framework integrating diverse machine-learning approaches and various available ET datasets to produce high-resolution, long-term ET estimates. We combine direct site observations and 13 different ET products that span remote sensing, machine-learning outputs, data fusion techniques, land surface models, and reanalysis datasets to create fused datasets. Our machine-learning framework integrated cutting-edge tools such as Automated Machine Learning (AutoML), Deep Neural Networks (DNN), Light Gradient Boosting Machine (LightGBM), and Random Forest (RF) algorithms to rigorously evaluate model efficacy. Our product exhibits a significantly improved spatiotemporal resolution (0.1 degree, daily) and extended temporal coverage (from 1950 to 2022) compared to existing datasets. In summary, this novel data integration framework overcomes previous ET data limitations through improved quality, spatiotemporal resolution, coverage, and advanced machine learning techniques. The resulting product will enable more accurate ET estimates for water, energy, and carbon cycle applications.