Simulation of extreme rainfall from CMIP5 in the Onkaparinga catchment using a generalized linear model

Simulation of extreme rainfall from CMIP5 in the Onkaparinga catchment using a generalized linear model

Due to the changes in global climate, the intensity, frequency and magnitude of heavy rainfall events are changing and this has been documented in many recent studies. Increasing trends in extreme rainfall directly affects infrastructure, agriculture as well as public and ecosystem health. So, projection of changes in extreme rainfall events is useful for policy making associated with climate change adaptation. General Circulation Models (GCMs) are the most important tool for climate change impact studies. But due to their coarse spatial resolution and their inability to capture local rainfall processes, GCMs cannot be used directly in hydrological impact studies. To bridge this gap downscaling has often been applied to transform GCM information to a finer resolution. There are broadly two types downscaling, namely dynamic and statistical methods. The latter is inexpensive and readily implementable compared to dynamic downscaling. Among several statistical downscaling techniques, generalized linear model (GLM) based downscaling techniques incorporate the spatial-temporal structure of rainfall. Because of this, GLMs have been used in several recent studies. However, application of this technique for downscaling of extreme rainfall events is relatively new. In this study, a GLM based multi-site downscaling technique has been applied using the GLIMCLIM software package for downscaling extreme rainfall in the Onkaparinga catchment at nine rainfall stations. A GLM was fitted to the observed rainfall conditioned to several large scale atmospheric and circulation variables from NCEP reanalysis data for the calibration period 1991 to 2005. This relation was used to simulate the daily rainfall for the validation period (1981 to 1990) using NCEP reanalysis and CSIRO MK3.6 historical data and for the future period 2041 to 2060 using RCP4.5 and RCP8.5 scenarios of CSIRO MK3.6. These daily rainfall series were used to estimate extreme rainfall indices such as consecutive dry days (CDD), rainfall events greater than 10mm/day (R10) and annual maximum daily (AM) rainfall. As far our we are aware, this is the first attempt where a GLM technique has been applied for downscaling extreme rainfall events using large scale data from CMIP5 GCMs at least for the Australian climate. The study reveals that the model performed reasonably well in reproducing the CDD and AM rainfall whereas it underestimated the R10 statistics in most of the months of the year when driven by NCEP reanalysis data. Although the R10 was underestimated, the trend and variability were simulated well. Performance of the model deteriorates when driven by CSIRO MK3.6 historical data. Simulation differences between NCEP reanalysis and MK3.6 can be attributed by the bias in the large scale atmospheric and circulation variables. AM daily rainfall was reasonably downscaled for NCEP reanalysis data over the period 1981 to 2005, whereas it was overestimated most of the time at all rainfall stations in the simulation driven by MK3.6. AM rainfall magnitudes for different Average Recurrence Intervals (ARIs) when fitted to Log Pearson Type III distribution were significantly larger for the period 2041 to 2060 under both RCP4.5 and RCP8.5 scenarios of MK3.6 compared to the observed data over the period 1981 to 2005. But the reduced accuracy in the simulation run using MK3.6 data may be due to bias in the large scale atmospheric and circulation variables relative to NCEP reanalysis. The study concludes that the GLM can be used to downscale extreme rainfall events. Also non-stationarity in relation to local rainfall and large scale climate variables is considered as a source of uncertainty in climate change impact studies. Most importantly adequate bias correction in the GCM data is essential before any projection is made.