Plant access to belowground moisture allows sustained evapotranspiration during drought

Subsurface water limitation is key to determine vegetation sensitivity to drought. There is evidence that water sources deeper than soil moisture exert a strong control on evapotranspiration (ET). However, these are impossible to observe directly at large scales and are not accounted for properly by land surface models. Here, we develop a method to study the impact of belowground water limitation on ET in progressing drought from ecosystem flux measurements. We train deep neural networks with 88,000 site-days of eddy-covariance data, meteorological data, multiple soil moisture datasets, and a remotely sensed greenness index. We derive a belowground water stress factor (fET) that isolates its ET reduction compared to atmospheric aridity, net radiation, air temperature and vegetation greenness. To investigate the sensitivity of ET impacts to progressing droughts, we regress fET against the cumulative water deficit (CWD), a normalized measure of whole-column water availability. We find that climate is the most important control on fET under drought, with much less variation induced by soil type or plant functional type. Evergreen needleleaf forests show the lowest sensitivity to belowground water stress, displaying almost negligible fET reductions up to a CWD of 300 mm. Savannahs and grasslands present an abrupt and progressive drop down to 10% fET after a CWD threshold of 50 mm. This variety of responses is not captured by a standard land surface model. We suggest this happens because models usually do not account for belowground water reservoirs. Our findings illuminate the understanding of vegetation sensitivity to drought and provide insight into modelling water stress effects.