Understanding the performance of downscaled global climate model precipitation data: a cautionary tale for watershed modeling applications

Abstract

There is a growing demand for accurate predictions of climate variables across various spatial and temporal scales, as the Earth’s climate has entered a period of rapid change that outpaces any historical experience of human civilizations (Kotamarthi et al., 2021; USGCRP, 2018). Climate models are the primary tools utilized for understanding climate patterns of the past and potential changes in future patterns as a result of decisions and choices made today (Edwards, 2011; Flato et al., 2013). At a large scale, global climate models (GCMs) can reasonably represent the interactions of climate variables throughout the climate system. However, global climate models lack the ability to reproduce climate variables, such as precipitation, at a scale that is applicable for specific hydrological regional or local assessments (Bastola and Misra, 2014; Cooney, 2012). Downscaling methods and techniques aim to refine the resolution of GCM outputs in order to provide more impactful and locally relevant climate projections for stakeholders and decision makers across regions, states, and cities. The intent of this research is to improve upon the current understanding and use of downscaled GCMs’ precipitation data for watershed modeling application and water resource management in the Blue River Watershed (BRW), an urbanizing watershed located within the Central United States Midwest region. Precipitation data was obtained from 18 downscaled GCMs under two different statistical downscaling methods: Localized Constructed Analogs (LOCA) and Multivariate Adaptive Constructed Analogs (MACA), for a total of 36 downscaled GCMs. Precipitation data from the 36 downscaled GCMs was statistically compared to precipitation data observed at six National Oceanic and Atmospheric Administration (NOAA) stations located throughout the BRW. From the statistical point comparison seven models were selected to assess spatial grid-surface comparisons of downscaled GCM data to the gridded spatial precipitation datasets obtained from Parameter-elevation Regressions on Independent Slopes Model (PRISM). The results from the point statistical analysis demonstrated that seven models performed better as compared to the whole group of 36 models assessed for the BRW. The performance of the seven downscaled GCMs were further assessed, by comparing modeled monthly data to observed monthly data across gridded surfaces, utilizing the Kolmogorov-Smirnov (KS) test and the SPAtial EFficiency metric. This further investigation emphasized that the selected downscaled GCM precipitation monthly comparisons had a poor performance in capturing the monthly magnitude and distribution of precipitation for the given selected models, downscaling methods, time period, and surface extents assessed throughout this study. Overall, this research stresses the importance of approaching the application of downscaled GCMs precipitation outputs, at the watershed scale, cautiously. Understanding the uncertainties and variability that remain in downscaled GCM precipitation outputs is critical for effectively incorporating into water resource management designs and decisions, as well as continually advancing the capabilities of climate scientist tools for application.