Assessment of MC&MCMC uncertainty analysis frameworks on SWAT model by focusing on future runoff prediction in a mountainous watershed via CMIP5 models

The river situation in a dry or semi-dry area is extremely affected by climate change and precipitation patterns. In this study, the impact of climate alteration on runoff in Kashafrood River Basin (KRB) in Iran was investigated using the Soil and Water Assessment Tool (SWAT) in historical and three future period times. The runoff was studied by MIROC-ESM and GFDL-ESM2G models as the outputs of general circulation models (GCMs) in the Coupled Model Intercomparison Project Phase 5 (CMIP5) by two representative concentration pathway (RCP) scenarios (RCP2.6 and RCP8.5). The DiffeRential Evolution Adaptive Metropolis (DREAM-ZS) was used to calibrate the hydrological model parameters in different sub-basins. Using DREAM-ZS algorithm, realistic values were obtained for the parameters related to runoff simulation in the SWAT model. In this area, results show that runoff in GFDL-ESM2G in both RCPs (2.6 and 8.5) in comparing future periods with the historical period is increased about 232–383% and in MIROC-ESM tends to increase around 87–292%. Furthermore, GFDL-ESM2G compared to MIROC-ESM in RCP2.6 (RCP8.5) in near, intermediate, and far future periods shows that the value of runoff increases 59.6% (23.0%), 100.2% (35.1%), and 42.5% (65.3%), respectively. 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/wcc.2019.122 ://iwaponline.com/jwcc/article-pdf/11/4/1811/830454/jwc0111811.pdf Armin Ahmadi Department of Civil Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran Amirhosein Aghakhani Afshar (corresponding author) Department of Water Engineering, Faculty of Civil Engineering, University of Tabriz, Tabriz, Iran E-mail: a.s.a.a.6269@gmail.com Vahid Nourani Center of Excellence in Hydroinformatics, Faculty of Civil Engineering, University of Tabriz, Tabriz, Iran and Near East University, Faculty of Civil and Environmental Engineering, Near East Boulevard, 99138, North Cyprus, via Mersin 10, Nicosia, Turkey Mohsen Pourreza-Bilondi Department of Water Engineering, College of Agriculture, University of Birjand, Birjand, Iran A. A. Besalatpour Inter3 GmbH Institut für Ressourcenmanagement, Otto-Suhr-Allee 59, 10585 Berlin, Germany


INTRODUCTION
The impact of climate change on the availability of water resources has been attracting the attention of researchers throughout the world (Faramarzi et al. ; Rodrigues et al. ; Zuo et al. ). In the context of precipitation frequency changes and climate change, numerous studies indicate that global warming leads to ecosystem degradation and water crisis (Falkenmark & Rockström ; Zuo et al. ) and exacerbated water scarcity in dry and semi-dry regions. About 600 million people live in areas with less than 500 m 3 water per capita, therefore enduring a severe shortage of water (Pereira et al. ).
Thus, there is a great deal of concern among researchers about the increasing threat of water scarcity not only in the region but also at global scale (Vörösmarty et al. ). As a result, in the global warming context, comprehensive assessment of water resources is a vital part of understanding the availability of water and improving water management towards maintainable, effective, and instrumental usage of scarce freshwater resources. Furthermore, climate change has a significant impact on water availability in river basins and there is an interaction between water resources and climate change (Faramarzi et al. ). Spatial and temporal variabilities of water resources in basins are severely impacted by climate change (Vörösmarty et al. ). Due to the high variation and low rate of precipitation in dry and semidry areas, many sectors can suffer direct or indirect effects on the economy and can lead to loss of people (Anyamba et al. ) and, most likely, effects on runoff variation   According to the literature review, the assessment of uncertainty analysis on the SWAT model with DREAM-ZS algorithm for predicting runoff with CMIP5 models in multisite calibration has not been reported yet. Therefore, to fill this gap, the current study was conducted to survey the uncertainty prediction capabilities of the DREAM-ZS algorithm to extract posterior ranges of parameters to predict runoff in future time periods.
In this research, the DREAM-ZS algorithm is applied as a formal Bayesian method for uncertainty analysis, calibration, and measurement of SWAT model parameters in the KRB.

MATERIALS AND METHODS
The study area   20, 10, and 20% threshold values for land-use, slope, and soil classes, respectively. The hydrometric stations are located at the outlet of the sub-basins and are selected in terms of area and topographic status to obtain a homogeneity between the sub-basins. By examining the overlapping of the maps, we concluded that by defining these thresholds, we can integrate small polygons with an area of less than 25 hectares, depending on the scale of the maps, into their larger adjacent polygons. Afterwards, a multiple slope option with four classes of 0-5, 5-10, 10-15 and >15% was selected for slope discretization (Afshar et al. ). Moreover, the modi- Finally, the monthly data set was used for calibration purposes, which related to an 11-year time span (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011).

DREAM-ZS algorithm
The DREAM algorithm has been employed as an effective where, g and q are length and number of the Markov chains, respectively; W is average q within sequences' variances, and B/g is variance between q sequence means.
Given that the values of posterior distribution are less than 1.2 for all of the parameters, it turns out to be stationary (Vrugt et al. a, b where Y and θ are the vectors of model output and unknown model parameter sets with d dimension, respectively. Parameter set (θ) in Bayesian framework can be calibrated by minimization of the errors e(θ), including distinction between corresponding observed output and the model prediction. However, the errors involved in the hydrologic models (data, structure, output, and parameter calibration errors) could be kept independent having PD with a constant variance and zero mean (Schoups & Vrugt ). For details about DREAM-ZS see Vrugt et al. (a).
In this study, a prior set of DREAM-ZS algorithm has been considered via the calibrated parameter set of calibration and uncertainty procedures (SWAT-CUP) and sequential uncertainty fitting program phase 2 (SUFI-2) process (  P-factor, NSE, and d-factor are in the range of 0-100%, À∞-1, and 0-∞, respectively. In the calibration process,    The best CMIP5 models with the highest fit to the observed data were selected on the following steps, after

RESULTS AND DISCUSSION
Since the observed data are very restricted, only the results related to the calibration data set are illustrated and the results of the validation data set are excluded.  (Table 4).
After these chains converged, the posterior parameter distributions of the SWAT model were produced with the last 20% of the parameters' set (i.e., 10,000 parameters' set). These plots are illustrated by Figure 5. Also, the best values (maximum likelihood) for all parameters are represented by (×) (see Table 5 Table 5. The uncertainty origins, such as U1 and U2, might be considered as the total (and parameter) uncertainty.
As can be seen in Figure 6, the most observation points (over 90%) of DREAM-ZS take place within 95% predictive uncertainty limits and P-factor demonstrates high performance of the DREAM-ZS algorithm. Also, the results illustrate that the total uncertainty (and   (Table 5) are not a very good representative of precipitation over the basin, because the map shows there are no rain gauges in some high and most low altitude parts of the study basin. Also, the KRB is a mountainous watershed and is exposed to high spatial and temporal variability in rainfall distribution that cannot be captured exactly using rain gauges in the basin. Therefore, such limitation could have contributed to some degree of simulation       Table 6 with regards to comparing runoff in future time periods and via both CMIP5 models with historical time.      A follow-up work to this study would be an investigation on water resource security (demand, availability, scarcity, reliability, and vulnerability) based on the blue and green water concepts by applying the mentioned models in this region.