Examining spatiotemporal trends of drought in the conterminous United States using self-organizing maps

Abstract

ABSTRACT Droughts are a natural, recurrent climate extreme that can inflict long-lasting devastation on natural ecosystems and socio-economic sectors. Unlike other natural hazards, drought onset is insidious and often affects a greater spatial extent over a prolonged temporal scale. In the United States the evolution of drought and its impacts are typically region-specific and intensified precipitation variability may obscure how drought may manifest. In this study, we examine the spatiotemporal trends of drought using self-organizing maps (SOM), competitive learning subset of artificial neural networks (ANN), requiring unsupervised training of inputs. We introduced monthly Palmer Drought Severity Index (PDSI) values to the SOM to identify existing clusters of wetting and drying patterns from 1895 to 2016. After training, we created cartographic visualizations of the SOM output and conducted a subsequent time-series analysis to link with the geographic patterns of drought. Over the last 40 years, precipitation intensified in the Northeast, Midwest, and upper Great Plains across several nodes. Across the majority of SOM patterns, we identified no significant changes of drying or wetting patterns over the last century for the greater part of the CONUS.