Evaluation of the Accuracy of CMIP6 Models based on Taylor Diagram for Simulating Precipitation in the Southern Part of the Aras River Basin

Introduction

Increasing surface temperature and change in precipitation patterns are the dominant phenomena in climate change and affect all parts of the water cycle. Precipitation is directly affected by the increase in global temperature, which increases the evapotranspiration rate, resulting in the concentration of water vapor in the atmosphere. The variety of different properties of precipitation, namely, intensity, frequency and temporal distribution in different areas is expected. Studies in Iran indicate a decrease in precipitation, increased temperature and water resources reduction under climate change. In order to illustrate climate change in future periods, 3D paired Atmospheric and Oceanic General Circulation Models (AOGCM) are the most prestigious tools. Given the dependence of climate change on the results of climatic models), it is mostly possible to achieve more reliable illustrations in the future using the climatic models presented in the new 6th report. The Coupled Model Intercomparison Project Phase 6 (CMIP6) continues the pattern of evolution and previous phases of the CMIP and includes new organized scenarios of global climatic modeling designed to identify different weather mechanisms. Compared to the CMIP5, the models in CMIP6 have generally improved better clarity and physical processes (Stouffer et al., 2017). Using a single AOGCM model to estimate temperature and precipitation changes in different parts of the world is a common approach. Research has shown that using a single model in this field may cause error and uncertainty in climate change forecasts. The reason for using higher -ranking models in terms of simulation skills is that models have different skills in different areas and courses. Uncertainty and bias vary from model to model and for specific variables. Therefore, direct use of model outputs is not recommended as it may lead to incorrect results. Therefore, the use of downscaling methods or bias correction for using climate data from GCM models is considered a necessary step. The results showed that temperature and precipitation would rise in the coming period. CMIP6 models predicted temperature and precipitation in the study area with a lower average error than CMIP5 models. The Aras basin has good conditions for agriculture and since agriculture requires water, and many dams have been built in the area that are the source of electricity generation in many areas, with climate change, precipitation patterns have changed and threaten the agriculture and water plant in the region. According to the above mentioned, it is important to study the precipitation of the Aras basin in future periods. This study is innovative in terms of downscaling with CMHYD software, using models with high horizontal resolution, and using four coupled 3D atmospheric and oceanic general circulation models to evaluate precipitation.



Data and Methodology

This study was aimed at evaluating the accuracy of selected models at 7 synoptic meteorological stations based on Taylor's diagram to simulate precipitation in the southern part of the Aras River Basin (Iran) over the past three decades. For this purpose, data from 4 AOGCM (MPI-ASM1-2-HR, CMCC-CM2-SR5, BCC-CSM2-MR and EC-Earth3-CC) were used by the CMIP6 series models. The historical period of 1985-2014 was considered. The raw output of the models downscaled by CMHyd software. To select the appropriate downscaling method, three methods were used: Linear Scaling, Power transformation, and Taylor diagram distribution mapping. The performance of the models at each station was evaluated by Taylor's chart. In this study, two groups of observational and model data were used daily for 3 decades. Daily precipitation data from 7 synoptic stations located in northwest Iran were obtained from the Iran Meteorological Organization (www.irimo.ir). The output of the CMIP6 was extracted from the https://esgf-node.llnl.gov/projects/cmip6/ for the time period 1985-2014 (historical period). In this study, initially, among the CMIP6 series models, those models that haddata with common historical periods and high horizontal spatial resolution were selected.After extracting the output of the 4 selected models, in order to reduce the uncertainty, 3 methods (Linear Scaling), Power Transformation and Distribution Mapping) were used for downscaling in the CMHyd software environment. The efficiency of each of the 3 downscaling methods was determined by drawing the Taylor diagram. The Taylor diagram simultaneously considers the skewness, standard deviation and root mean square indices. In this diagram, the observed data is specified as a reference point on the horizontal axis and the angular dimension indicates the correlation between the observed and simulated values. The standard deviation values are plotted as concentric circles with respect to the center of the circle and the RMSE values are plotted as concentric circles with respect to the reference point.The output of GCM models cannot be used directly due to their large scale. To overcome the problem of low spatial resolution, downscaling methods are used. The CMHyd software was used to downscale the output of general circulation models. This software was developed for hydrological modeling by Rathjens et al, 2016 at Purdue University, USA, in the Python environment. The CMHyd uses eight bias correction methods in a separate process for precipitation and temperature. Out of the eight methods, five methods are specific to precipitation. Fast execution and the ability to select different options are the advantages of this software.This software requires three types of data, including observation data, historical climate model data, and scenario data (future) climate models.The downscaling process is performed in five steps, which are: entering observation variables (in text form), selecting a bias correction method, entering model data in the historical and scenario periods (in text form or NetCdf), processing (including checking data and performing downscaling), and outputting results in both numerical and graphical forms.Historical simulations are useful for assessing the accuracy of models. Historical model periods are an important tool for determining the consistency and sensitivity of climate models to observational data and controlling the uncertainty of these models (Eyring et al., 2016). To determine the accuracy of each of the 4 models in this study at 7 stations in northwest Iran, after removing bias, Taylor plots were used. After determining the best models for each station, the raw and downscaled output of the best models for each station was verified with the Kling-Gupta Efficiency (KGE) statistical index. The KGE index is a composite index that is able to combine several statistical indices such as mean, standard deviation, and data correlation with each other and increase the accuracy of model selection based on their ability to simulate the historical period. In other words, when several statistical indicators are used separately in this field, it will be difficult to make a final decision without using the KGE index.



Results

The Taylor diagram of Jolfa station for 3 models in the base period, based on the output of 3 methods: Linear Scaling, Power Transformation, and Distribution Mapping, showed that the Linear Scaling method has high accuracy and less error than the other two methods. In the case of low values of the correlation coefficient, it is necessary to note that with a complex and volatile precipitation variable, it is not possible to expect a high value of the correlation coefficient; while the correlation coefficient in temperature studies shows high values. Based on the evaluation performed by the Taylor diagram, among the three bias correction methods, the Linear Scaling method was selected and the downscaling of the raw output of the models in this study was performed using the above method in the CMHyd software.After downscaling, the precipitation output of the 7 synoptic meteorological stations selected in this study in the historical period (1985-2014) was verified by 4 models with Taylor diagrams. According to the Taylor diagram drawn at each synoptic meteorological station, the BCC model had the best output at all stations in the study area with small values of standard deviation and also root mean square error, of course, at Jolfa station, the first place belongs to the CMCC model by a small difference. Regarding small values for correlation, it should be noted that the uncertainty of the precipitation variable is much higher than the temperature variable and the correlation coefficient cannot be expected to have significant values. According to the drawn Taylor diagrams, the standard deviation is in the range of 0.5 to 1 and the RMSE is in the range of 1 to 1.5 and the Pearson correlation coefficient is in the range of 0 to 0.1 at all stations. After validating the 4 models used in this study using Taylor diagrams, the best models were identified by synoptic station in the southern part of the Aras basin. In all selected stations except Jolfa station, the best model among the 4 models was the BCC model. According to the calculations, the downscaling performed by the Linear Scaling method has appropriately optimized the output of the models in all stations of the study area and reduced their errors.

Conclusion

the results indicated that the Linear Scaling fine-tuning method had better capabilities based on the verification test conducted among the other 2 downscaling methods of this study (Power transformation and Distribution mapping). The evaluation results with the KGE measure for raw and downscaled output values of all stations indicate the appropriate performance of the Linear Scaling fine-tuning method. The calculations showed that the top model in all selected stations in the study area is the BCC model and the weakest model for simulation of the southern part of the Aras River Basin, the MPI model. The results showed that the raw output of the models had a lot of error and could not be used directly. The results also showed that the Linear Scaling downscaling method has a good ability to optimize the output of GCM models in the study area.