Machine learning-based investigation of forest evapotranspiration, net ecosystem productivity, water use efficiency and their climate controls at meteorological station level

Evapotranspiration (ET), net ecosystem productivity (NEP), and ecosystem water use efficiency (EWUE) of forests are changing due to climate change. Traditional models using coarse-scale climate reanalysis data fail to capture local meteorological and hydrological conditions accurately. This study combines in situ meteorological observations, remote sensing, and advanced datasets (forest age, rooting depth, soil moisture) to estimate ET, NEP, and EWUE at forest meteorological stations via machine learning. About 60.6 % of stations showed a decrease in NEP from 2003 to 2010 to 2011–2019, while 63.9 % showed an increase in ET, and 58.9 % showed a decrease in EWUE. NEP and EWUE significantly declined in forests older than 60 years, with younger forests exhibiting higher NEP. EWUE in different forest types is driven by varying mechanisms, with DBF sites influenced by VPD and ENF sites by RSDN. EWUE of regions with inconsistent VPD data between site and reanalysis, such as northwestern North America, showed divergences from previous reanalysis-based studies but aligned more with atmospheric inversion findings. Slight summer VPD increases boosted NEP, especially in high-latitude areas, while early spring phenology and increased spring VPD reduced summer water availability. Incorporating more site-specific observations, such as plant traits, could enhance understanding of climate-plant-ecosystem relationships. This study underscores the potential of meteorological station-level data to improve forest carbon and water flux dynamics understanding, aiding forest management for carbon neutrality and climate adaptation.