A framework for assessing the drivers and impacts of drought events: the contemporary drought in the western and central United States

Abstract We develop a framework for assessing the drivers and impacts of droughts, built upon a Markov Random Field and machine learning-based drought identification algorithm that operates simultaneously in space and time. The method uses a precipitation threshold for drought, while considering the drought state of neighboring grid points and identifies contiguous and distinct droughts that propagate through space and time. Importantly, this method can identify droughts of any scale, from a single grid point to those encompassing many thousands. We apply it to North American precipitation from observations and a multi-model ensemble of 67 historical simulations to produce a repository of 25,156 identified droughts. The framework uses an observed drought for comparison, and we choose the 2011-2014 drought in the western and central United States, which is among the most severe and persistent in recorded history. As the spatiotemporal characteristics of the simulated droughts become more like the observed drought, we quantify if their local-scale impacts (evaporation, leaf area index, soil moisture, and runoff) and large-scale drivers (atmospheric circulation, sea surface temperature, and modes of climate variability) become predictable. Our findings suggest that ecological impacts are not predictable even when simulated droughts closely match the spatiotemporal characteristics of the observed drought. The drought drivers are also not predictable, with similar droughts occurring under a range of atmosphere-ocean conditions. These results suggest that the drivers and impacts of even the most persistent and severe droughts have limited predictability, although additional work is needed to quantify the role of structural uncertainty and better understand the real-world applicability of climate model-based results.