MGML-ET: integrating TSEB-derived features with machine learning for large-scale evapotranspiration mapping

Evapotranspiration describes how water moves from land to the atmosphere through soil evaporation and plant transpiration. It strongly influences water availability, ecosystem health, and climate, yet it is difficult to estimate accurately over large areas, especially where landscapes are complex or observations are sparse. This challenge is becoming more serious as climate change increases water stress and environmental variability. In this study, we developed a new approach to estimate daily evapotranspiration by combining satellite observations, weather data, and information from a physical energy balance model with machine learning. Instead of forcing physical equations into the model, we used physically meaningful estimates to guide the learning process. This allows the model to remain flexible while improving its physical realism. The results show that the new approach produces more reliable and stable evapotranspiration estimates than existing methods, especially across different land-cover types. We also found widespread increases in evapotranspiration linked to large-scale vegetation restoration projects. This framework provides a practical way to monitor how ecosystems and water resources respond to climate change and land management, offering useful information for sustainable water planning and ecological restoration.