Enhancing Daily Evapotranspiration Estimates in Germany Using Multi-Source Data and Machine Learning Models

Evapotranspiration (ET) is the process describing the transfer of water from the land surface into the atmosphere. It includes evaporation from soil and plant surfaces, as well as transpiration through plant stomata. ET represents a central link between the terrestrial water cycle, energy cycle, and carbon cycle. Thus, an accurate estimation of ET is essential for understanding the landscape water budget and biomass production, as well as for improving agricultural water management and irrigation strategies. Additionally, more accurate ET values improve models of carbon-water interactions in land ecosystems and promote sustainable water use.

This study aims to identify an appropriate model for estimating daily ET across a west-east climate and land-use gradient in Germany, providing an effective method that accurately reflects ET variability. The main objectives of this study are to (i) estimate ET using a combination of machine learning, physics-based, and hybrid models; (ii) evaluate the performance, efficiency, and sensitivity of these models by comparing estimated ET with ET observations; and (iii) use the most accurate model to predict daily ET variability along the climate and land use gradient in Germany.

To achieve this, remote sensing data from Sentinel-2 and Landsat-8, as well as meteorological data from the German Weather Service (DWD) and the ERA-5 reanalysis, were used for model training. To assess the models' performance, eddy-covariance data from the Integrated Carbon Observation System (ICOS) and ET products from the Moderate Resolution Imaging Spectroradiometer (MODIS) for the years 2017 to 2024 were used.

We used five different approaches to estimate daily ET, including deep learning (DL), machine learning, hybrid models, and physical models. Specifically, we employed the Optical Trapezoid Model (OPTRAM), Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Random Forest (RF), and TabTransformer. We evaluated the accuracy of each approach using ICOS ET observations.

The results indicated that DL models generally performed better than RF and OPTRAM-ET models in the study area. Among all the experiments, the ANN achieved the best performance, with a root mean square error of 0.6 and a correlation coefficient of 0.91. Additionally, we observed significant variations in modeling performance across different ecosystem types. In grassland, ET estimates showed the highest accuracy, whereas in cropland ecosystems, the greatest deviations were observed.