Using machine learning to quantify multi-scale soil moisture controls on water and carbon fluxes at the land surface

<p><span lang="EN-US">Knowledge of fluxes of water vapor and carbon at the land surface are paramount to our understanding of the Earth system. Large-scale network initiatives such as the Fluxnet allow us to better understand the environmental controls on the evapotranspiration and gross primary productivity. An important aspect of such initiatives is that its large number of sites allow for localized knowledge to be upscaled to a region or even globally. This can be either done by employing physics-based global land models or empirically, via data-driven approaches. Particularly, we have seen a significant increase of data-driven approaches with the use of machine learning techniques more recently. Here, we use a similar structure employed in the FLUXCOM initiative to focus particularly on the role of soil moisture information in predicting evapotranspiration and gross primary productivity at several flux sites encompassing a wide range of hydroclimates and biomes around the globe. Our analyses employ a machine learning method to a predictive model of evapotranspiration and gross primary productivity, while focusing primarily on how changes in the way soil moisture is incorporated into the methodology affects such predictions. First, we evaluate the predictive power of this model when soil moisture is directly estimated via observations against more indirect estimates via bucket-type models. Secondly, we evaluate the role of the spatial resolution of different soil moisture estimates in predicting both fluxes. We do this by using three sets of direct estimates covering distinct spatial footprints co-located at all flux sites: (1) point-scale time-domain reflectometers, (2) field-scale cosmic-ray neutron sensors, and (3) regional-scale satellite remote sensing products. In this talk, we summarize which hydroclimatic regions benefit from having direct estimate of soil moisture for evapotranspiration and gross primary productivity, while also providing some insights on the possible role of spatial scale mismatches between the fluxes and soil moisture.</span></p>