Climate model bias correction for nonstationary conditions

In most of the climate change impact assessment studies, climate model bias is considered to be stationary between the control and scenario periods. Few methods are found in the literature that addresses the issue of nonstationarity in correcting the bias. To overcome the shortcomings reported in these approaches, three new methods of bias correction (NBC_μ, NBC_σ, and NBC_bs) are presented. The methods are improvised versions of previous techniques relying on distribution mapping. The methods are tested using split sample approach over 50-year historical period for nine climate stations in Ontario, using six regional climate models. The average bias reduction improvement by new methods, in mean daily and monthly precipitation, was found to be 73.9%, 74.3%, and 77.4%, respectively, higher than that obtained by the previous methods (eQM 67.7% and CNCDFm_NP 64.1%). Thus, the methods are found to be more effective in accounting for nonstationarity in the model bias.

Striking stationarity of large-scale climate model bias patterns under strong climate change

Because all climate models exhibit biases, their use for assessing future climate change requires implicitly assuming or explicitly postulating that the biases are stationary or vary predictably. This hypothesis, however, has not been, and cannot be, tested directly. This work shows that under very large climate change the bias patterns of key climate variables exhibit a striking degree of stationarity. Using only correlation with a model’s preindustrial bias pattern, a model’s 4xCO2bias pattern is objectively and correctly identified among a large model ensemble in almost all cases. This outcome would be exceedingly improbable if bias patterns were independent of climate state. A similar result is also found for bias patterns in two historical periods. This provides compelling and heretofore missing justification for using such models to quantify climate perturbation patterns and for selecting well-performing models for regional downscaling. Furthermore, it opens the way to extending bias corrections to perturbed states, substantially broadening the range of justified applications of climate models.