Validation of downscaling models for changed climate conditions: case study of southwestern Australia

Abstract

Statistical downscaling of general circulation models (GCMs) and limited area models (LAMs) has been promoted as a method for simulating regional-to point-scale precipitation under changed climate conditions. However, several studies have shown that downscaled precipitation is either insensitive to changes in climatic forcing, or inconsistent with the broad-scale changes indicated by the host GCM(s). This has been recently attributed to the omission of the effect that changes in atmospheric moisture content have on precipitation. We describe validation of a nonhomogeneous hidden Markov model (NHMM) for changed climate conditions and apply it to a network of 30 daily precipitation stations in southwestern Australia. NHMMs fitted to 1 × CO 2 LAM data were validated by assessing their performance in predicting 2 × CO 2 LAM precipitation. The inclusion of 850 hPa dew point temperature depression, a predictor reflecting relative (rather than absolute) atmospheric moisture content, was found to be crucial to successful performance of the NHMM under 2 × CO 2 conditions. The NHMM validated for the LAM data was fitted to the historical 30 station network and then used to downscale the 2 × CO 2 LAM atmospheric data, producing plausible predictions of station precipitation under 2 × CO 2 conditions. Our results highlight that the validation of a statistical downscaling technique for present day conditions does not necessarily imply legitimacy for changed climate conditions. Thus statistical downscaling studies that have not attempted to determine the plausibility of their predictions for the changed climate conditions should be viewed with caution.

Simulation of seasonal precipitation and raindays over Greece: a statistical downscaling technique based on artificial neural networks (ANNs)

“…Even though one of the most important assumptions of any statistical downscaling technique is the stationarity of the parameters used (Wilby, 1997), there is no assurance that the relationship of the predictors and the predictants will not change in the future (Maheras et al, 2004). In this sense, the downscaled results should be interpreted cautiously (Charles et al, 1999).…”

Linking climate change modelling to impacts studies: recent advances in downscaling techniques for hydrological modelling

“…Karl et al, 1990;Wilby and Wigley, 1997;Murphy, 2000;Beckmann and Buishand, 2002); particularly as it may be an important predictor under a changed climate. Indeed, the inclusion of moisture variables as predictors can lead to convergence in the results of statistical and dynamical approaches (Charles et al, 1999), with the inclusion of GCM precipitation as a predictor also improving downscaling skill (Salathé, 2003;Widmann et al, 2003). Cavazos and Hewitson (2005) have performed the most comprehensive assessment of predictor variables to date, assessing 29 NCEP reanalysis variables using an artificial neural network (ANN) downscaling method in 15 locations.…”

Stochastic simulation of rainfall in the semi‐arid Limpopo basin, Botswana

“…The introduction of the extra unobservable stochastic state is the primary difference between this and other multi-site methods, potentially allowing a more flexible structure for linking atmospheric variables to rainfall temporal and spatial patterns . The model has been applied widely in Australian catchments over a range of climates for different purposes, including water resources assessment and climate change analysis (Charles et al, 1999b(Charles et al, , 2004(Charles et al, , 2007Mehrotra et al, 2004;Robertson et al, 2006), and has been found capable of reproducing the magnitude and inter-annual variability of rainfall during dry and wet periods. As with most multi-site models, NHMMs may need a large number of parameters to satisfactorily reproduce observed rainfall, potentially creating high uncertainty in simulations.…”