Characterising climate model bias: An insight into bias nonstationarity and spatial dependence in hydroclimatic simulations

Abstract

Biases in climate model simulations are one of the biggest challenges in climate change impact assessment studies. The systematic biases, introduced in the simulations from the imperfect representation of climate models, need to be addressed prior to their use for hydrologic modelling. This thesis investigates various aspects of climate model bias and advances the important field of to an improvement of bias correction. The common assumption of stationary temporal biases is tested for precipitation and temperature simulations across Australia. It is found that there is considerable bias non-stationarity at decadal and multi-decadal time scales of GCM simulations. Difference in biases between positive and negative phases of the IPO along the east coast of Australia, where IPO impacts tend to be the strongest challenges the inherent assumption in most bias correction approaches. Despite this finding the nature of the bias non-stationarity suggests that it will be difficult to modify existing bias correction approaches to account for non-stationary biases. Bias propagation in space is an issue while transferring information from coarse to fine scales through statistical downscaling. Although, the statistical downscaling method BCSD corrects biases in the distribution at GCM scale, at finer spatial scales biases are re-introduced as the observed fine scale spatial variability is not properly represented by mean anomaly that is used in the method. A rank-based approach that models the anomalies in rank-space can address this limitation and improves the distributions of the climate variables, particularly for low and high precipitation amounts. Further extending the consideration of spatial bias, an ICA based two-step bias correction method is proposed to correct time series at multiple locations conjointly. The first step uses ICA to transforms the data to a set of statistically independent univariate time series for bias correction. The spatially corrected data is then bias corrected at the grid scale which provides climate simulations that have greater equivalency in space and time with observational data. The method is shown to also correct biases that propagate in time and space through dynamic downscaling. The spatially corrected precipitation time series shows skill for high rainfall amounts at the catchment scale.