Downscaling the Relationship between the Precipitation and Runoff of the Rood Zard Basin in the Climate Change Context

Studying and identifying the climate variabilities occurring in different regions, may give insight toward possible future climate variabilities.  Using available climate models as well as downscaling, is a way to recognize the possible variabilities of climate components of future. In the present study, the precipitation and runoff of Rood Zard Basin were downscaled and simulated for the time period of 2006-2100. For this purpose, the RCP output scenarios of the CanESM2 model were utilized for 1975-2005. For downscaling the precipitation and runoff of Rood Zard Basin, the daily precipitation data of Baghmalek and the runoff data of the Mashin station and artificial neural network method were used. The mean sea-level pressure, the geopotential height at 500 hPa, and the mean temperature at ground level were all selected as the predictive variables, using correlation and partial correlation calculations, as well as the backward variable elimination method. The verification of the design was carried out by the RMSE and R2 indexes. Finally, the network architecture was selected through the Bayesian Regularization algorithm along with three hidden layers as the optimal network. The results show that annual precipitation have decrease trends in future 95 years. revealed that the precipitation increased in the hot months of the year and decreased in the cold months. In other words, the increase of local rainfalls due to the temperature rise is most probable in future periods. The runoff would decrease in the cold months and increase in the warm months regardless of the temperature and vegetation impact. Climate change is the main phenomenon affecting the climate and the human environment as well as environmental phenomena (such as droughts and wetness years, water resources, sea level changes, temperature alterations, changes in the behavior of climate elements, and many other phenomena). Investigating many phenomena of the past decades revealed that the planet earth's climate is changing. Compared to the previous time periods, the  results of the previous studies indicated that the climate variablity trend has become faster in the past 150 years. To fully understand the climate, all the units involved in its formation should be evaluated simultaneously. For this purpose, models may be helpful to some extent. Modeling is the process of creating a model that can provide the structure and function of systems. One of these methods is GCM in which the climate is simulated. These models are developed based on different climate scenarios aiming to simulate the impact of greenhouse gases on the earth's climate. Moreover, they are able to simulate and predict the future climate of the earth. These models create various time series of climate variables with relatively large networking. However, they are not suitable for direct use in the studies relating to the local climate variability. Thus, researchers have designed suitable downscaling methods to gain the climate data on a local scale. One of these methods is the statistical downscaling.

In the present study, the precipitation and runoff of the Rood Zard Basin are downscaled based on the RCP climate scenarios. RCPs are new emission stimulant scenarios which are used as the input of CMIP5 climate models and are based on the fifth report of IPCC. Scenarios are important parts of climate simulations that allow the researchers to study the long-term outcomes of the current decisions. In the RCP scenarios, 26 atmospheric parameters were considered for future simulations. Each of these has a relatively high connection to environmental elements. The selection of the most optimal parameter for expressing the relationship between weather conditions and the environmental characteristics depends on the type of environmental parameters. To select the appropriate parameters, the correlation and partial correlation calculations and the Backward Variable Elimination methods were applied.For downscaling, the BOX_019X_44Y data were acquired from the Environment website of Canada. The data were analyzed through calculating the correlation coefficients, partial correlation and also the Backward Variable Elimination method. The results revealed that 3 variables including the Mean Sea Level Pressure, the geopotential height at 500 hPa, and the mean temperature at ground level had an acceptable correlation with the precipitation at the Baghmalek station and omitting other variables created a lower missing variance.Downscaling was carried out based on the artificial neural network model with the Bayesian Regularization algorithm. Artificial neural networks are the patterns for processing data which are produced by imitating the neural network of the human brain. In recent decades, this method has been recognized as a useful and reliable tool for modeling complex maps existing between different variables. Artificial neural networks are able to pick up a system’s hidden behavior through available data. Each network has three layers: the input layer, the hidden layer, and the output layer. The input layer is, in fact, a layer used for producing the data given to the network as an input. The output layer includes values that are simulated by the network. The hidden layer is the place of analyzing the data. Unusually, the number of chosen neurons in this layer is obtained through trial and error.

In order to downscale neural network using the output RCP scenarios of the CanESM2 model, the daily precipitation data in the Baghmalek station during a time period of 30 years (1975-2005) were chosen as the base statistical period. After the selection of atmospheric high-scale variables, these variables were introduced into the neural network as input. The precipitation was considered as the target and the network was designed using algorithms and numerous hidden layers. Finally, the network designed with the Bayesian regularization and 3 hidden layers were chosen as the optimal network. As mentioned earlier, the artificial neural network was used for downscaling. Moreover, the daily precipitation data were simulated for the statistical period of 2006-2100. Linear regression was applied for simulating the runoff for the aforementioned period. The daily runoff, as well, was estimated for this period. The results demonstrated that the estimated monthly precipitation rate from November to December in the future 95-year period has decreased. Likewise, the simulated precipitation rates from January to November were higher than the monthly precipitation rates in the base period. Therefore, it can be concluded that the precipitation decreased in the cold months and increased in the hot months.  Additionally, the runoff in the base period from January to May was less than the observed runoff and it was more than the observed runoff from June to December. This was due to the fact that only precipitation was used as an independent variable for modeling; whilst, the runoff was affected by other factors such as springs water in addition to the rainfall. From November to May, the estimated monthly rates of runoff for the next 95 years were reduced. Moreover, from November to October, the simulated runoff rates were more than the monthly runoff rates in the base period. Accordingly, it can be concluded that the runoff decreased in the cold season and increased in the hot season, as well. The increase in the precipitation and runoff rates in the hot season could be due to the rise in the local rainfalls. In other words, an increase in the local rainfalls due to global warming was probable in future periods.