Assessing impacts of future climate change on hydrological processes in an urbanizing watershed with a multimodel approach

The sensitivity of hydrological processes to the changed environment is of great concern. The integrated impacts of climate change and urbanization in the future have been assessed in a watershed in Northwest China through a multimodel approach based on the combined application of Generalized Watershed Loading Functions, the Long Ashton Research Station Weather Generator, and the Land Change Modeler. The results showed that both climate change and urbanization would lead to more watershed streamflow, and their combination would have synergistic effects on additional increases. In addition, there would be different seasonal distributions of streamflow with a greater proportion of runoff. These study results are helpful in supporting projects and/or decisionmaking processes for managers by providing more insights into the regional hydrological changes affected by climate change and urbanization. The proposed methodology of the combined multimodel approach may be applicable in other areas with similar conditions.


INTRODUCTION
Water is critical for human communities (Milly et al. ; Vorosmarty et al. ). The responses of watershed hydrological processes to the changing environment are of vital importance for regional security and sustainable development which are of great concern (Tan & Gan ; Wada et al. ). Regional climate change is one of the most significant natural influences on local hydrological properties (Oki & Kanae ; Kristvik et al. ). As the source and driving force of streamflow, precipitation is the key factor in hydrological processes. In addition, the air tem- Previous studies have found various responses for regional water resources in the changes of climate and land use in different areas. The increase in precipitation would lead to more potential streamflow, but the higher temperature would lead to less streamflow resulting from the increase of evapotranspiration. The more impervious surface from urbanized land-use cover would lead to more water transfers from a runoff route that forms streamflow more easily but is consumed by evapotranspiration. In addition, the combined effects of climate and land-use change on hydrological processes were intricate and of great concern in recent studies (Marhaento et al. ; Wang & Kalin ). Generally, the response of the hydrological process would be different in different areas, and it is important to obtain estimations for various watersheds.
The main objective of the study is to address the changes of watershed hydrological variables under various individual and combined scenarios of climate change and urbanization with proper model approaches in a watershed in northwest China.
As a source tributary of the Yellow River, the Jing River was used as a study case in this paper. Given that it is critical for regional water resource management, it is important to estimate its hydrological response to future changed environments with proper model applications. However, the available data for the study area were limited, and establishing a valid modeling scheme is of great significance. In this study, we proposed an integrated multimodel approach to estimate the responses of hydrological processes to modeled future climate change and land-use conversion scenarios for the Jing River watershed. The hydrological model of Generalized Watershed Loading Function (GWLF) was employed as the modeling tool for the watershed hydrological estimation here, because its modest data requirement was able to be satisfied with the existing data from the study watershed. The Land Change Modeler (LCM) for ArcGIS was operated with its multilayer perceptron neural network function for future land-use predictions. Considering the spatial accuracy of the GWLF model, future climate changes were estimated in overall watershed scale.
As a cultivated and woodland-dominated watershed with a small proportion of the urban area, there would be an assumption of stationarity for watershed climate status, and the weather generator model of the Long Ashton Research Station Weather Generator (LARS-WG) was eligible for the statistical downscaling analysis (Salvi et al. ).
In addition, the spatial non-uniform and local instability caused by urban heat island effect from an increased urban area were ignored (Shastri et al. ), and various synthetic sequences of daily weather variables generated by the LARS-WG based on different GCM outputs were used in GWLF for the hydrological response assessment.
All three models have been widely applied in their own domains, and benefit from their coherences in scope, the potential of multimodel linkages is concerned and tested for a reliable assessment, to the status of future watershed hydrological processes considering the synergistic effects of climate change and urbanization.

Study area and data source
This study was conducted in the Jing River watershed, located in northwestern China. The Jing River extends from the eastern side of Liupan Mountain and drains into the Wei River, which is the largest tributary of the Yellow River. The portion of the Jing River watershed located above the Jing-Chuan Hydrologic Gauge Station was used as the study area. The area of the study watershed is approximately 3,145 km 2 , with a mean daily temperature of 9.03 C and a mean annual precipitation of 503.31 mm. It is a multiple land-use watershed with upland forest and grass in upstream areas and considerable cultivated land in downstream areas along the river. Ping-Liang City is located in the study watershed. It is surrounded by a number of small villages and experiences critical pressure from water resource shortages. The area of Ping-Liang City increased by over 150%, with significant population growth in the past few decades, indicating a significant demand for response estimations and predictions of future water resources to support local management. The main geographical and environmental attributes of the study watershed are shown in Figure 1, and the sources of the original data used in this study are summarized in Table 1.

Application of GWLF
The GWLF model was employed to model watershed hydrological processes (Haith & Shoemaker ). It was used here due to its moderate data requirement, which can be   1956-1963, 1971-1990, and 2006-2014  were used in the calibration process, while the records of 1956-1963 and 2006-2014 were reserved for verification.
The sensitive transport parameters determined by an advanced sensitivity analysis were calibrated by a GLUE Bayesian analysis. The prior distributions for each parameter were set as the default ranges from the GWLF manual, and 100,000 groups of parameter sets were sampled followed by the same number of model iterations to build a reliable likelihood function distribution based on the contrast of modeled and observed data for the calibration period. The Nash-Sutcliff coefficient (R 2 NS ) was used to measure model accuracy, and a value of 0.80 for R 2 NS was set as the cutoff threshold for posterior sampling to determine the parameters' distributions as well as the uncertainty of the results. The calibrated GWLF could then be used to estimate the responses of hydrological processes to various changes in environmental factors.

Application of the LCM
The LCM is a software extension for ArcGIS developed by Clark Labs (Eastman et al. ). Being a powerful tool, the LCM has been widely used for the assessment and prediction of land cover change and its implications (Fuller In this study, there were 13 land-use types that existed in the study area, and the land-use maps in 2000 and 2010 were used as the earlier and later land cover images, respectively. The change analysis was operated by the LCM, and five significant transitions that were greater than 1 km 2 were of concern, including cultivated land to shrubbery lands, middle coverage grassland, cities and towns, and rural residential land, as well as other forestland to cities and towns. Six potential driving or explanatory variable maps were employed, including elevation, slope, distance to road, distance to city, and distance to village. These maps were evaluated for each transition between two land-use types by using the MLP neural network to select the best combination and create related submodels and transition potential maps, based on which future land cover maps could be predicted. The land-use map of 2015 was used for validation, and the land-use maps of 2050 and 2080 were modeled with a Markov Chain process to represent land cover conversion linked to the GWLF model to estimate future hydrological processes.

Application of the LARS-WG
The LARS-WG stochastic weather generator was used for the downscaling analysis to obtain a future synthetic weather time series (Semenov & Barrow ). It uses a series of semi-empirical distributions to describe weather factors, the parameters of which are calibrated and validated with long-term observational weather data records. The LARS-WG has a significant capability to reproduce the stationarity weather condition with the implicit assumption that history would repeat itself and the statistical characteristics based on historical weather records would be still valid for the future. In addition, by updating the model para- To quantify the changes of climatic factors under various AR4 scenarios, six general circulation models (GCMs) were employed, including HADCM3, GFCM21, INCM3, IPCM4, MPEH5, and NCCCSM. The outputs of each GCM had already been embedded into the LARS-WG in advance. An ensemble approach is adopted by using mean values of multi-GCMs to avoid uncertainty from using one single GCM, based on which the calibrated LARS-WG parameters were updated to generate future synthetic weather data series. Sixty years of synthetic daily weather data were generated for each scenario in one future period to represent the predicted climate conditions in the study area, which could be further used as input weather data for the GWLF to estimate the hydrological response.
A flow chart was provided in Figure 2 to illustrate the process of this study.

Hydrological model
The time series of the observed and modeled monthly streamflow during the research period are illustrated in

Land-use change model
The results of the MLP neural network in the LCM for each land-use cover transition are listed in Table 3. All accuracy rates were higher than 85%, and all the skill measures were higher than 0.75, indicating that the selected variable maps could drive and/or explain the land cover changes,   (Table 4). In  3: Distance to rural residential land, which is dynamic to be updated during calculation.
4: Distance to cities and towns, which is dynamic to be updated during calculation. 5: Distance to river, which is dynamic to be updated during calculation.

Hydrological responses
Hydrological process in future climate changes The changes in watershed hydrological processes under future climate conditions were estimated. Seven synthetic series of 60 years of daily weather data that indicated the current and future status were input into the GWLF, while the land-use cover from 2010 was constantly used for all  streamflow was similar to that of the mean value but had a greater increase in intensity. The mean rate of increase of the extreme annual streamflow for B1 scenarios in the 2080s would be 12.1%, higher than the 10.7% from mean value-based statistics. In addition, the increase of extreme annual streamflow for A1B scenarios in the 2050s would be more than 15%. These results implied a greater flood risk under future climate conditions, which should be of great concern for local water resource management. The In addition, the changes in climate conditions would also convert the initial monthly apportionment of critical hydrological factors (Figure 7). As a result of the rise in air