Characterizing uncertainty in sea level rise from 900 future groundwater depletion scenarios

Sea level change is driven by numerous processes which vary across scales both spatially and temporally. Data-driven model abstractions for the natural and anthropogenic processes governing sea level are used to make future sea level projections through the remainder of the 21st century and beyond, both globally and by region. In addition to the primary drivers of sea level rise (mass loss of terrestrial ice and thermal expansion of seawater), changes in terrestrial water storage, namely groundwater extraction and reservoir impoundment, have also been identified as contributors. Uncertainty surrounding future projections of groundwater depletion and its implications on global and relative sea level have thus far been inadequately characterized and underexplored, and rely on a small number of studies based on simple methods with little or no scenario dependence or spatial variability. We use a 900-member large ensemble of basin-level groundwater depletion simulations through 2100 from a global integrated assessment model to explore a wide outcome space of groundwater futures and characterize their critical drivers from among six systematically varied inputs: socioeconomics (SSPs), climate forcing (RCPs), climate model (GCM), groundwater availability, surface water storage, and hydrological model. The large ensemble of simulations enables a more robust discussion of uncertainty and scenario dependence than previously available. We find the median global mean sea level (GMSL) rise by 2100 due to groundwater depletion in our model ensemble to be 83 (17-205) mm. We also find that the greatest concentrations of groundwater depletions take place in the Western US, the Nile Basin, the Middle East, and Central and South Asia, though the correlation between basins can vary widely. This basin-level dataset also enables sea level fingerprinting to assess the spatially variable effects of groundwater depletion on relative sea level (RSL). Uncertainty in this fingerprinting can then be compared with the uncertainty bounds traced by the large ensemble of model runs.