Coupled mechanistic ecohydrological modelling and machine learning can provide robust estimates of global ET and GPP

Global estimates of evapotranspiration (ET) and carbon sequestration (GPP) are of outmost importance for hydrological, ecological and climate modelling. However, high quality ground data availability is limited in many places in the world limiting our skill in quantifying ET and GPP at the global scales. Detailed models resolving the coupled water and carbon cycles can provide a mechanistic understanding of the links between climate and vegetation dynamics. Such models commonly require data not available in sparsely gauged locations in the world making their applicability limited for global scale applications.

In this study we use the model output simulated by the land surface and terrestrial ecosystem model T&C for 79 well monitored stations in the world to train two different machine learning algorithms in order to derive global estimates of ET and GPP. The sites cover most biomes spanning from arid shrublands to tropical forests. The machine learning algorithms include Artificial neural networks (ANN) and Random Forests (RF). The external co-variates of the machine learning algorithms include remotely sensed leaf area index (MODIS - LAI), solar induced fluorescence (GOME2 - SIF), soil moisture (SMAP), and global datasets of climate (CRU), vegetation cover and plant physiological traits, and soil hydraulic properties. Such an approach can yield robust models of global estimates of ET and GPP that can continuously improve with real time data assimilation.

Various previously reported remote sensing estimates for ET and GPP were used as benchmarks. ANN-simulated ET(GPP) shows Nash-Sutcliffe Efficiencies (NSE) of 0.72(0.8) on average compared to the benchmarks. RF ET(GPP) estimates show a NSE of 0.72(0.67) on average.