An intercomparison of statistical downscaling methods used for water resource assessments in the United States

Abstract

Information relevant for most hydrologic applications cannot be obtained directly from the native-scale outputs of climate models. As a result the climate model output must be downscaled, often using statistical methods. The plethora of statistical downscaling methods requires end-users to make a selection. This work is intended to provide end-users with aid in making an informed selection. We assess four commonly used statistical downscaling methods: daily and monthly disaggregated-to-daily Bias Corrected Spatial Disaggregation (BCSDd, BCSDm), Asynchronous Regression (AR), and Bias Corrected Constructed Analog (BCCA) as applied to a continental-scale domain and a regional domain (BCCAr). These methods are applied to the NCEP/NCAR Reanalysis, as a surrogate for a climate model, to downscale precipitation to a 12 km gridded observation data set. Skill is evaluated by comparing precipitation at daily, monthly, and annual temporal resolutions at individual grid cells and at aggregated scales. BCSDd and the BCCA methods overestimate wet day fraction, and underestimate extreme events. The AR method reproduces extreme events and wet day fraction well at the grid-cell scale, but over (under) estimates extreme events (wet day fraction) at aggregated scales. BCSDm reproduces extreme events and wet day fractions well at all space and time scales, but is limited to rescaling current weather patterns. In addition, we analyze the choice of calibration data set by looking at both a 12 km and a 6 km observational data set; the 6 km observed data set has more wet days and smaller extreme events than the 12 km product, the opposite of expected scaling.

Multivariate quantile mapping bias correction: an N-dimensional probability density function transform for climate model simulations of multiple variables

“…Coarsened CanRCM4 outputs are more spatially coherent (median I = 0.38) than the WFDEI observations (median I = 0.18). Univariate QDM maintains this unrealistic spatial coherence (median I = 0.33), which is consistent with results reported by Gutmann et al (2014). Conversely, MBCr leads to precipitation fields that are less coherent than observed (median I = 0.07).…”

“…Overall, spatiotemporal statistics from MBCn match observed values more closely than the other methods. Hence, it may be feasible to use MBCn directly in downscaling applications, a practice that has been questioned for univariate quantile mapping (Maraun 2013;Gutmann et al 2014).…”

“…For example, Abatzoglou and Brown (2012), Stoner et al (2013), Ahmed et al (2013) each applied quantile mapping to daily climate model data that had been interpolated to a high-resolution observational grid. However, Maraun (2013) and Gutmann et al (2014) demonstrated that this approach-applying a univariate bias correction algorithm to interpolated climate data at individual grid points-can lead to fields with unrealistic spatial structure, especially if the variable being downscaled operates on spatial scales that are substantially finer than the climate model grid. This is particularly true for precipitation.…”

Mapping (dis)agreement in hydrologic projections

“…Two other important sources of uncertainty are the statistical downscaling technique of the GCM output and the internal climate variability. Gutmann et al (2014) showed that different downscaling techniques each have their advantages and disadvantages. The technique employed here, bias-corrected spatial disaggregation (Wood et al, 2004), tends to overestimate the wet-day fraction and underestimate extreme events, which both can influence the hydrologic response of the catchment.…”

“…The representation of underlying (physical) principles of hydrological processes in the hydrologic model can thus have a profound effect on the results and conclusion of a study. Although uncertainty in hydrologic projections has already been discussed and investigated in the literature, studies usually focus on one source of uncertainty (Gutmann et al, 2014) or a limited number of catchments (Vidal et al, 2016;Addor et al, 2014;Dobler et al, 2012). Here, we investigate three sources of uncertainty (GCM forcing, hydrologic parameters, hydrologic model structure) in hydrologic projections for 605 basins throughout the contiguous US over a wide range of climates, in order to reveal spatial patterns in uncertainty.…”