Nonparametric methods for modeling GCM and scenario uncertainty in drought assessment Subimal Ghosh 1 and P. P. Mujumdar 1 Received 17 July 2006; revised 26 January 2007; accepted 22 February 2007; published 6 July 2007. [1] Hydrologic implications of global climate change are usually assessed by downscaling appropriate predictors simulated by general circulation models (GCMs). Results from GCM simulations are subjected to a number of uncertainties due to incomplete knowledge about the underlying geophysical processes of global change (GCM uncertainties) and due to uncertain future scenarios (scenario uncertainties). With a relatively small number of GCMs available and a finite number of scenarios simulated by them, uncertainties in the hydrologic impacts at a smaller spatial scale become particularly pronounced. In this paper, a methodology is developed to address such uncertainties for a specific problem of drought impact assessment with results from GCM simulations. Samples of a drought indicator are generated with downscaled precipitation from available GCMs and scenarios. Since it is very unlikely that such small samples resulting from GCM scenarios fit a known parametric distribution, nonparametric methods such as kernel density estimation and orthonormal series methods are used to determine the probability distribution function (PDF) of the drought indicator. Principal component analysis, fuzzy clustering, and statistical regression are used for downscaling the mean sea level pressure (MSLP) output from the GCMs to precipitation at a smaller spatial scale. Reanalysis data from the National Center for Environmental Prediction (NCEP) are used in relating precipitation with MSLP. The information generated through the PDF of the drought indicator in a future year may be used in long-term planning decisions. The methodology is demonstrated with a case study of the drought-prone Orissa meteorological subdivision in India. Citation: Ghosh, S., and P. P. Mujumdar (2007), Nonparametric methods for modeling GCM and scenario uncertainty in drought assessment, Water Resour. Res., 43, W07405, doi:10.1029/2006WR005351. 1. Introduction [2] General circulation models (GCMs) are tools designed to simulate time series of climate variables for the world, accounting for the effects of the concentration of greenhouse gases in the atmosphere [Prudhomme et al., 2003]. Coupled with projections of CO 2 emission rates, they produce climate scenarios that can be described as ‘‘pertinent, plausible representations of the future emissions of greenhouse gases and with the understanding of the effect of increased atmo- spheric concentration of the gases on global climate’’ [IPCC- TGCIA, 1999]. They are currently the most credible tools available for simulating the response of the global climate system to increasing greenhouse gas concentrations, and they provide estimates of climate variables (for example, air temperature, precipitation, wind speed, pressure, etc.) on a global scale. GCMs might capture large-scale circulation patterns and correctly model smoothly varying fields such as surface pressure, but it is extremely unlikely that these models properly reproduce nonsmooth fields such as preci- pitation [Hughes and Guttorp, 1994]. Additionally, the spatial scale on which a GCM can operate [for example, 3.75° longitude  3.75° latitude for coupled global climate model (CGCM2)] is very coarse for hydrologic applications [Prudhomme et al., 2003]. Downscaling is therefore neces- sary to model the hydrologic variables (for example, preci- pitation) at a smaller scale based on larger-scale GCM outputs. Dynamic downscaling uses complex algorithms at a fine-grid scale (typically of the order of 50  50 km) describing atmospheric processes nested within the GCM outputs [Jones et al., 1995] to result typically in limited-area models or regional climate models (RCM), whereas statis- tical downscaling produces future scenarios based on sta- tistical relationships between large-scale climate features (for example, circulation pattern) and hydrologic variables [Wilby et al., 1998]. A major assumption in the statistical downscaling is that there are certain physical relationships underlying the statistical relationships developed, and these physical relationships hold regardless of whether the model simulation is a control (stationary) experiment or an exper- iment incorporating changed climate [Easterling, 1999]. Compared with dynamic downscaling, statistical down- scaling has the advantage of being computationally simple and easily adjustable to new areas. The method generally requires very few parameters, and this makes it attractive for many hydrologic applications [Wilby et al., 2000]. 1 Department of Civil Engineering, Indian Institute of Science, Bangalore, India. Copyright 2007 by the American Geophysical Union. 0043-1397/07/2006WR005351$09.00 W07405 WATER RESOURCES RESEARCH, VOL. 43, W07405, doi:10.1029/2006WR005351, 2007 Click Here for Full Articl e 1 of 19