Impact of Climate Change on Drought Characteristics and Its Uncertainty              

                                    Hung-Wei Tseng

This study aims to assess the impact of climate change on a water resources system of southern Taiwan. For this purpose, firstly, the study proposed a two-step variable elimination procedure for monthly precipitation spatial downscaling based on Random Forests (RF) for finding out the possible impact on hydrology. The proposed procedure not only can reduce the variable dimension but also preserve the model performance simultaneously. Two variable importance (VI) estimates (i.e., error-based and impurity-based estimates) provided by RF were applied to judge whether a variable should be kept or not. The minimized important variables were regarded as the predictors for precipitation downscaling. The precipitation projections under various emission scenarios were made according to the latest climate change experiments of five general circulation models. The downscaling results show that the future precipitation may increase in the wet season (from May to October) but decrease in the dry season (from November to April). Secondly, based on these downscaling results, a weather generator was applied to work with the downscaling information for synthesizing daily weather data under various climate change scenarios. Since a stochastic process is involved in weather generation, the study deals with this uncertainty by using ensemble mean approach. Then, the daily weather data were input to a hydrological model for deriving the scenario runoffs under various climate change scenarios. Based on these projection results, a large prime variation in scenario runoffs occurs in the wet season but a slight variation in the dry season. In terms of quantity, the scenario projections indicate scenario runoffs probably increase in the wet season. Lastly, the scenario runoffs were routed through the reservoir operation model to collect all failure events (when the water supply does not meet the water demand) of the water resources system. The study proposed a composite drought index, water shortage index (WSI), to quantify these failure events of the water resources system. The results show that there is a 26.6% chance (four out of the fifteen sets) that climate change impact could lead to more severe droughts in the study site in the near-term period. Most WSI values show that the impact of climate change has a positive effect on then water resources system due to a rise in runoffs in the wet season. However, since all of the future scenario projections were made based on the five selected GCMs from the CMIP5 in which more than fifty GCMs are included, the results of this study may therefore partly represent the possible future. It is suggested including more GCMs for assessing the uncertainty among GCMs.

Considering sedimentation is a serious problem for the Tsengwen Reservoir, the additional effect of sedimentation to the impact of future drought events to the study site was also considered. Under the combined impact of climate change and the effect of sedimentation which could decrease the storage capacity of the Tsengwen Reservoir by 12%, more severe drought impacts will be expected in five of the fifteen cases. In the worst case, WSI turns out to be 4.979 which is more than triple that of the baseline value of 1.597 when the total amount of water shortage is projected to be 1.9 times that of the baseline case. Moreover, the water resources system in the worst case scenario will not be able to provide enough water around two thirds of the operation time, and it has a lower probability to return to a satisfactory state once a drought event of that magnitude has occurred. The reduction of the reservoir capacity by sedimentation aggravates the potential impact of drought events, leading to more serious water deficits, than if only the hydrologic impact of climate change on the study site is considered.

Keywords: spatial statistical downscaling, Random Forest, variable elimination, drought indices, monotonic behavior.