Selection of CMIP5 general circulation model outputs of precipitation for peninsular Malaysia

Reduction of uncertainty in climate change projections is a major challenge in impact assessment and adaptation planning. General circulation models (GCMs) along with projection scenarios are the major sources of uncertainty in climate change projections. Therefore, the selection of appropriate GCMs for a region can significantly reduce uncertainty in climate projections. In this study, 20 GCMs were statistically evaluated in replicating the spatial pattern of monsoon propagation towards Peninsular Malaysia at annual and seasonal time frames against the 20th Century Reanalysis dataset. The performance evaluation metrics of the GCMs for different time frames were compromised using a state-of-art multi-criteria decision-making approach, compromise programming, for the selection of GCMs. Finally, the selected GCMs were interpolated to 0.25 × 0.25 spatial resolution and biascorrected using the Asian Precipitation – Highly-Resolved Observational Integration Towards Evaluation (APHRODITE) rainfall as reference data. The results revealed the better performance of BCC-CSM1-1 and HadGEM2-ES in replicating the historical rainfall in Peninsular Malaysia. The biascorrected projections of selected GCMs revealed a large variation of the mean, standard deviation and 95% percentile of daily rainfall in the study area for two futures, 2020–2059 and 2060–2099 compared to base climate. This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/). doi: 10.2166/nh.2020.154 ://iwaponline.com/hr/article-pdf/51/4/781/730910/nh0510781.pdf Saleem A. Salman Mohamed Salem Nashwan Tarmizi Ismail (corresponding author) Shamsuddin Shahid School of Civil Engineering, Faculty of Engineering, Universiti Teknologi Malaysia, 81310 Johor Bahru, Malaysia E-mail: tarmiziismail@utm.my Mohamed Salem Nashwan Department of Construction and Building Engineering, College of Engineering and Technology, Arab Academy for Science, Technology and Maritime Transport (AASTMT), 2033 Elhorria, Heliopolis, Cairo, Egypt


INTRODUCTION
The uncertainties in climate projections heavily influence the quantification of impacts (Deser et al. ). A slight variation in the changes in climate projection can significantly change the return period of hydrological disasters such as flood, GCMs are generally selected according to their ability to simulate historical climate (Lutz et al. ). Mostly, time series of monthly or annual observed and GCM simulated climate are compared to assess the performance of GCMs.
Selection of GCMs by conventional approach provides emphasis on their ability to simulate temporal variability in rainfall (Ahmed et al. ; Noor et al. b). The ability of GCMs in simulating spatial variability in climate is often ignored although it has similar importance. Besides, GCMs are also selected based on their capability to simulate largescale ocean-atmospheric phenomena responsible for climate variation of a region. For example, the ability of GCMs to simulate monsoon is very important to show its capability to project rainfall in a monsoon-dominated rainfall region.   On the other hand, APHRODITE provides daily estimates of rainfall on a 0.25 × 0.25 spatial grid over land Out of more than 45 GCMs available in CMIP5, 20 GCMs were initially selected in this study as they have daily historical data  and future projections (2020-2099) for the four RCPs (i.e., 2.6, 4.5, 6.0 and 8.5) to cover a wide range of possible future changes. The GCM name, modelling centre and spatial resolution are provided in Table 1.

METHODOLOGY Procedure
The procedure of selecting and bias-correcting a subset of GCMs for reduction of uncertainty in rainfall projection for Peninsular Malaysia is outlined below: 1. The CMIP5 GCMs' rainfall simulations were interpolated onto a common resolution of 2.00 × 2.00 using bilinear interpolation.  6. The biases in the selected GCMs' simulations were corrected using linear scaling method by using APHRODITE as a reference dataset for the historical period.

The performances of
7. The bias-correction factors were then applied to correct the bias in the projected estimates of the selected GCMs for the four RCPs.
The bias-corrected rainfall estimated by the selected GCMs were used to show the spatial and temporal changes in the mean, standard deviation and 95th percentile of the daily rainfall during the near and far futures (2020-2059 and 2060-2099, respectively) compared to a base period . Details of the methods are discussed below.

Statistical evaluation of GCM outputs
Six statistical metrics were used to evaluate the performance of the GCMs in simulating spatial variability of rainfall at three time frames (i.e., annual, NEM and SEM), separately, against the 20CR rainfall for the historical period 1961-2005. First, GCM data were interpolated to 20CR spatial resolution using the bilinear interpolation method. The bilinear interpolation method was used as it can smoothly interpolate   (7) where z is the number of evaluation metrics used, x j is the normalized value of metric j obtained for a certain GCM, x j * is the normalized ideal value of the metric j, and p is the parameter (1 for linear, and 2 for squared Euclidean distance measure). In the present study, the linear measure was used and, therefore, the value of p was considered equal to In this study, the same rule of thumb was adopted.   and R 2 (i.e., 75%, 0.45, 0.66 and 0.52, respectively). However, the GCMs underestimated the NEM rainfall by a median of 9.00% bias, as shown in Figure 5(a). The median rSD at the annual time frame was closer to the optimal one than at NEM and SWM.

Compromise programming and models ranking
In order to decide and select the near-optimum GCMs based on the statistical metrics results, CP was used to estimate the distance of each GCM from the ideal point. In this study, the nearest to the optimal value of a statistical matric (i.e., the lowest NRMSE%, the highest NSE, md and R 2 , the nearest rSD to 1, and the nearest Pbias to 0) was used as the ideal value of the corresponding metric. L cp ¼ j0:08 À 0:00j þ j0:73 À 0:64j þ j0:47 À 0:59j þ j0:66 À 0:72j þ j0:52 À 0:67j þ j0:91 À 0:97j ¼ 0:55 (9) The were ranked ascendingly as shown in Table 3.
The same procedure was carried out for ranking GCMs in simulating rainfall for the SWM and annual time frames. Table 4 shows the final rank of GCMs in simulating spatial variability of NEM, SWM and annual rainfall in Peninsular Malaysia. As shown in the table, CMCC-CMS was ranked last for NEM, while it achieved the first position in simulating both SWM and annual rainfalls. This means, that the CMCC-CMS was able to simulate the annual and SWM rainfall spatial pattern very well, but completely failed in simulating NEM rainfall. Therefore, following the rule of thumb suggested by Ahmed et al. (), CMCC-CMS was discarded. A similar procedure was followed for all the GCMs and only the GCMs that achieved a rank between 1 and 10 for the three time frames were selected for the ensemble. This means that the final subset of GCMs should be able to simulate the annual and both monsoon rainfalls adequately. As shown in Table 4, two GCMs -BCC-CSM1-1 and HadGEM2-ESshowed acceptable performance for all the three time frames and were selected in the final subset of GCMs, although their individual ranks were not the highest. Therefore, those two GCMs were selected for future projection of rainfall in Peninsular Malaysia.

Bias correction of the selected GCMs
The selected GCMs, BCC-CSM1-1 and HadGEM2-ES, were bias-corrected using linear scaling method for the study area. Figure 6 shows scatter plots of the monthly raw and bias-corrected rainfall (presented by cross and circle However, the bias-corrected rainfall was found to underestimate the higher rainfall and overestimate the lower rainfall  values. This is very common in bias correction of GCMs using any method (

CONCLUSION
An attempt has been made to reduce uncertainty in the projection of rainfall of Peninsular Malaysia through the