Statistical downscaling and disaggregation for supporting regional climate change impact studies

Abstract

A warmer climate may affect the frequency and severity of weather extremes, such as heavy rainfalls, hurricanes and heatwaves. Based on the records around the world, the numbers of observed extreme events have presented increasing tendencies over the past decades. The Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report points out that the temperature would be continuously increasing in this century. This implies that some disasters (e.g. flood and drought) which are caused by weather extremes could become more frequent. The Southeast Asia is vulnerable to the impact of climate change. Especially in the urban areas, the flash flood has become one of major disasters caused by heavy rainfall. It is thus critical to develop flexible and applicable approaches to investigate the climate change impact on local regions. The General Circulation Models (GCMs) are the powerful tools to simulate either current or future climate conditions. But the bias and resolution problems have limited their applications for some specific regions like Southeast Asia. The dynamical and statistical downscaling are the two basic approaches to help bridge the gaps between GCMs and local weather information. Compared with dynamical approaches, the statistical ones are computationally cheap and easily applicable to many different regions. The objective of this PhD study is to develop and apply statistical downscaling and disaggregation methods for supporting hydrological and climate change impact studies. It covers three major components including development of novel statistical downscaling tools, applications of combined statistical downscaling and disaggregation methods, and assessment of climate change impact on hydrological processes. Firstly, two novel methods of statistical downscaling approaches were developed. One was a regression-based method which included the K nearest neighbor (KNN) and Bayesian neural network (BNN) models. In this method, KNN was used for classifying the dry/wet day and rainfall typing based on rainfall intensity. BNN was vii applied for prediction of rainfall amount. The other one was a semi-empirical stochastic weather generator which was mainly based on the Markov chain, semi-empirical random generator, and KNN. In detail, a four-state first-order Markov chain was employed to estimate day status which includes dry-day, wet-day with low-intensity rain, wet-day with moderate-intensity rain and wet-day with high-intensity rain based on the mean areal rainfall; then, three semi-empirical distributions were used to fit the three wet categories of rainfalls; finally, the KNN method was used in spatial disaggregation to generate daily rainfalls at multiple stations. A common idea of the two models was to use classification technique to describe different rainfall magnitudes based on rainfall intensity. The study results demonstrated that classification was helpful to enhance characterization of the convective rainfalls in the tropical region in the downscaling processes. Secondly, a number of com...