Unraveling the Hydrology of the Glacierized Kaidu Basin by Integrating Multisource Data in the Tianshan Mountains, Northwestern China
Abstract
Understanding the water balance, especially as it relates to the distribution of runoff components, is crucial for water resource management and coping with the impacts of climate change. However, hydrological processes are poorly known in mountainous regions due to data scarcity and the complex dynamics of snow and glaciers. This study aims to provide a quantitative comparison of gridded precipitation products in the Tianshan Mountains, located in Central Asia and in order to further understand the mountain hydrology and distribution of runoff components in the glacierized Kaidu Basin. We found that gridded precipitation products are affected by inconsistent biases based on a spatiotemporal comparison with the nearest weather stations and should be evaluated with caution before using them as boundary conditions in hydrological modeling. Although uncertainties remain in this data‐scarce basin, driven by field survey data and bias‐corrected gridded data sets (ERA‐Interim and APHRODITE), the water balance and distribution of runoff components can be plausibly quantified based on the distributed hydrological model (J2000). We further examined parameter sensitivity and uncertainty with respect to both simulated streamflow and different runoff components based on an ensemble of simulations. This study demonstrated the possibility of integrating gridded products in hydrological modeling. The methodology used can be important for model applications and design in other data‐scarce mountainous regions. The model‐based simulation quantified the water balance and how the water resources are partitioned throughout the year in Tianshan Mountain basins, although the uncertainties present in this study result in important limitations.
1 Introduction
The assessment of hydrological response to climate change is a vital research field in the Tianshan Mountains and many other mountain regions (Biskop et al., 2016; Chen et al., 2016b; Doris et al., 2016; Ragettli et al., 2016; Viviroli et al., 2007). Water resources supplied from the Tianshan Mountains (known as the “Water Tower of Central Asia”) are of great importance for downstream rivers, residents, irrigation agriculture, and ecosystems (Chen, 2014; Hagg et al., 2007; Shen & Chen, 2010; Sorg et al., 2012). Climate‐driven changes have a significant influence on hydrological regimes in snow‐fed and glacier‐fed basins (Barnett et al., 2005); the Tianshan Mountains were found to be critical in forcing large‐scale circulation changes (Baldwin & Vecchi, 2016). It is therefore an important research task to improve our understanding of mountain hydrological processes in high‐elevation catchments in the Tianshan Mountains.
High‐altitude basins are vulnerable to the combined impacts of warming temperature, precipitation variability, and changes in snow and glacier dynamics in the Tianshan Mountains (Duethmann et al., 2015; Sun et al., 2015). Temperature shows a positive trend (Chen et al., 2006; Shi et al., 2007) and snow cover extent and glacier sizes have been decreasing (Yao et al., 2004; Ye et al., 2005; Yong et al., 2007). However, streamflow has been increasing (Chen et al., 2016b; Tao et al., 2011). The water cycle is likely to become more unstable (Shen & Chen, 2010). Additionally, rising temperatures have changed glacier mass balances and snowfall fraction, which leads to less snow accumulation and alters the role of meltwater in the regional water balance (Chen et al., 2016a). In the Kaidu Basin on the southern slope of the Tianshan Mountains, climate‐driven changes not only increase the volume of streamflow (Chen et al., 2009, 2016b; Deng et al., 2015; Shen et al., 2018; Tao et al., 2011; Wang et al., 2013), but also the seasonal variability of streamflow (Liu et al., 2011) and glacier melt (Liu et al., 2006). Snowmelt runoff timing is expected to shift toward earlier dates due to spring temperature increases (Liu et al., 2011; Shen et al., 2018). Snow and glacier meltwater play an important role in seasonal patterns of streamflow in this glacierized catchment. Although statistical analyses of observational data have revealed changes of hydrological regimes due to climate change (Chen et al., 2013), it is still necessary to quantify the contributions of different runoff components to streamflow in glacierized basins in order to gain a better understanding of the ongoing changes.
Hydrological modeling is widely used to understand hydrological processes at the basin scale. However, the application of hydrological models in glacierized basins is complicated either by inadequate observational data (Biskop et al., 2016; Dou et al., 2011; Fang et al., 2015; Ragettli et al., 2013; Zhang et al., 2007) or the limited knowledge of snow and glacier melt dynamics (Chen et al., 2016b; Sun et al., 2015), which are the main obstacles for the application of hydrological models in the Tianshan Mountains and downstream basins. For instance, insufficient and coarse data sets posed a major challenge in the optimization of the grid‐based Variable Infiltration Capacity (VIC) model in the Tarim Basin (Liu et al., 2010). In other earlier studies near our study region, glacier ablation was not fully represented by the Hydrologiska Byråns Vattenbalansavdelning (HBV) model in the glacierized Urumqi Basin (Sun et al., 2015), or neglected by the MIKE SHE model in the Tarim Basin (Liu et al., 2013). Furthermore, the SWAT model was enhanced to include glacier melt processes for simulating glacier retreat and its response to climate change in the Manas Basin in the Tianshan Mountains (Luo et al., 2013), yet solar radiation and topographic factors were excluded in the degree‐day factors. In the glacierized Kaidu Basin, the lack of observational data and the omission of glacier melt processes have been two major sources of uncertainty in earlier studies (Chen et al., 2016b; Dou et al., 2011; Liu et al., 2012; Xu et al., 2016a; Zhang et al., 2007, 2016). Efforts have been made to correct model parameters and apply hybrid models (Fang et al., 2015; Xu et al., 2016b); in a comparative study, a physically based model (MIKE SHE) performed better than a lumped conceptual model for spatially representing climate variability (Liu et al., 2011). In existing studies, glacier melt was excluded (Chen et al., 2016b; Dou et al., 2011; Xu et al., 2016a; Zhang et al., 2007, 2016), and the distribution of runoff components was not addressed (Liu et al., 2011; Sun et al., 2015; Xu et al., 2016a). To our knowledge, high (spatial and temporal) resolution physically based modeling has not been conducted in the glacierized Kaidu Basin, where it would be important to characterize the water balance and determine the relative contributions of different runoff components.
Gridded data sets offer the potential to fill data gaps, and have been widely used over the years for representing climatic patterns in mountainous regions (Biskop et al., 2016; Immerzeel et al., 2015; Shea et al., 2015). Remote sensing data can partly overcome data scarcity in hydrological modeling, yet their limitations in accuracy and temporal resolution need to be addressed in detail (Liu et al., 2012). Besides, gridded data quality is affected by complex topography and elevation effects in the semiarid Tianshan Mountains (Wang et al., 2015). The Global Precipitation Climatology Centre (GPCC V7) was previously utilized to investigate variations of annual precipitation in Central Asia (Hu et al., 2017), yet different precipitation products show large discrepancies in mountainous areas. The gridded data sets show different uncertainties and biases and therefore cannot be used directly without quality assessment and comparison, especially in mountainous regions (Gao et al., 2012; Wang et al., 2015). Taken together, the reliability of gridded data sets has not sufficiently been addressed and compared in the Tianshan Mountains, and their suitability for driving hydrological models remains unexplored.
Field data have the potential to further improve bias corrections and hydrological model performance (Immerzeel et al., 2014; Ragettli et al., 2015). Nevertheless, field data in the Tianshan Mountains are very limited due to inaccessible terrain. For instance, the temperature lapse rate is crucial for representing mountain temperature and simulating snowmelt runoff in high‐elevation basins (Deng et al., 2015; Immerzeel et al., 2014; Li & Williams, 2008; Lundquist & Cayan, 2007). According to our field research, temperature lapse rates vary spatially and seasonally in the Tianshan Mountains (Shen et al., 2016), which has not been addressed in previous hydrological modeling studies (Dou et al., 2011; Xu et al., 2016a; Zhang et al., 2007, 2016). Moreover, vegetation structure, soil type, geology, and morphological features were not considered in sufficient detail in previous studies (Liu et al., 2011; Sun et al., 2015; Xu et al., 2016b).
In the light of the mentioned limitations, the main objectives of this study are to unravel the glacierized mountain hydrology and characterize the distribution of runoff components in order to better cope with the variability of water resources. To achieve these goals, the applicability of gridded climate data sets was evaluated, and suitable data sets were subsequently corrected based on field data. Driven by multiple input data sets, a fully distributed hydrological model was applied. Moreover, uncertainty and sensitivity related to model parameters and equifinality were assessed. This study sheds light on hydrological processes in a data‐scarce glacierized basin. The results—although somewhat preliminary due to data scarcity—are of great importance for better understanding of the vulnerability of water resources in Central Asia and other mountainous regions with limited data availability.
2 Study Area and Data Sets
2.1 Study Area
This study focuses on the Kaidu Basin, which is located on the central southern slopes of the Tianshan Mountains (42°14′N–43°21′N, 82°58′E–86°05′E) in northwestern China (Figure 1). The basin drains an area of 18,649 km2 with a mean elevation of 3,100 m above sea level (a.s.l.) above the Dashankou gauge station (Figure 1a and Table 1). The Kaidu river originates from the Tianshan Mountains, flows through the Bayinbuluke grassland and finally arrives at Lake Bosten (Figure 1a), for which it is the largest tributary, accounting for about 87% of its mean annual inflow (Chen, 2014). Water released from Lake Bosten is used for irrigation and is of critical importance for ecosystems located further downstream.
(a) Location of the Kaidu Basin in Central Asia, elevation and location of observed and HOBO logger stations. Overview of the available geospatial data: (b) land cover classification, (c) soil types, and (d) lithology.
| Category | Name | Longitude (°E) | Latitude (°N) |
Elevation (m a.s.l.) |
Tmean (°C) |
Precipitation (mm) |
Time period |
|---|---|---|---|---|---|---|---|
| Long‐term weather stations | Bayinbuluke | 84.02 | 43.03 | 2458 | −4.25 | 272 | 1961–2011 |
| Baluntai | 86.11 | 42.73 | 1739 | 7.12 | 208 | 1961–2011 | |
| Kuche | 82.51 | 41.72 | 1082 | 11.30 | 70 | 1961–2011 | |
| Yanqi | 86.31 | 42.08 | 1055 | 8.58 | 76 | 1961–2011 | |
| Luntai | 84.11 | 41.78 | 976 | 11.20 | 65 | 1961–2011 | |
| Kuerle | 86.01 | 41.75 | 932 | 11.78 | 55 | 1961–2011 | |
|
Gauging stations |
Dashankou | 85.74 | 42.25 | 1340 | 1972–2008 | ||
|
HOBO temperature stations |
H1 | 83.93 | 42.71 | 2428 | −2.37 | Sep 2014 to Aug 2015 | |
| H2 | 83.71 | 42.89 | 2470 | −4.52 | Sep 2014 to Aug 2015 | ||
| H3 | 84.56 | 42.77 | 2483 | −2.06 | Sep 2014 to Aug 2015 | ||
| H4 | 84.17 | 42.94 | 2525 | 0.36 | Sep 2014 to Aug 2015 | ||
| H5 | 83.69 | 42.69 | 2663 | −2.65 | Sep 2014 to Aug 2015 | ||
| H6 | 83.33 | 42.92 | 2791 | −0.90 | Sep 2014 to Aug 2015 | ||
| H7 | 85.50 | 43.14 | 2986 | −0.96 | Sep 2014 to Aug 2015 | ||
| H8 | 85.51 | 43.19 | 3427 | −2.67 | Sep 2014 to Aug 2015 | ||
| H9 | 85.53 | 43.22 | 3771 | −4.35 | Sep 2014 to Aug 2015 |
Situated in northwestern China, the Kaidu Basin has a continental semiarid climate. Mean annual precipitation and temperature at the Bayinbuluke weather station are 272 mm and −4.25°C, respectively (1961–2011) (Table 1). Precipitation generally increases with altitude in the mountainous regions, and temperature experiences distinct spatial‐temporal variation (Chen, 2014; Shen et al., 2016). Precipitation and temperature are highly variable. More than 60% of the average annual precipitation at the Bayinbuluke occurs in the summer months. Year‐to‐year precipitation variation is highest in summer and lowest in winter; the opposite holds true for temperature. Seasonal mean temperatures at the Bayinbuluke are −1.5, 10.1, 2.1, and −20.3°C in spring, summer, autumn, and winter, respectively.
The Kaidu Basin is a rainfall‐fed and snow/glacier meltwater‐fed basin with very little human land use. The main land cover types are grassland and barren land (62% and 30%, respectively) (Figure 1b). Annual streamflow at the Dashankou gauging station is approximately 189 mm/yr (1972–2008). The observed increase of streamflow in the Kaidu Basin could be either due to the uptrend of precipitation or an increase of snow and glacier meltwater (Shi et al., 2007). Snow accumulates from November to March and is released in the spring and summer. Thus, spring streamflow is dominated by snowmelt, and glacier meltwater contributes to summer streamflow, yet precipitation is the main source of discharge in summer (Deng et al., 2015; Fu et al., 2013). Base flow also contributes a vast proportion (41%) of water throughout the year (Chen et al., 2009). However, the distribution of runoff components is insufficiently studied.
2.2 Data Sets
Daily stream discharge data from the Dashankou gauging station were collected from the Hydrology and Water Resources Bureau of Xinjiang. Daily precipitation and temperature data were obtained from the China Meteorological Data Service Center (CMDC) (https://data.cma.cn/en). HOBO Pro V2 (U23‐001) temperature loggers (HOBO) were furthermore installed in the field about 2 m above the ground surface between 2428 and 3771 m a.s.l. in the Kaidu Basin. Location and summary information of hydrometeorological stations are provided in Figure 1 and Table 1.
Several global and regional gridded precipitation products were evaluated in this study (Table 2). They include interpolated data: Asian Precipitation‐Highly‐Resolved Observational Data Integration Towards Evaluation (APHRODITE) (Yatagai et al., 2012) and Climatic Research Unit (CRU) (Harris et al., 2014); reanalysis data: ERA‐Interim (conducted by European Centre for Medium‐Range Weather Forecasts, ECMWF) (Dee et al., 2011), Modern‐Era Retrospective Analysis for Research and Applications, version 2 (MERRA‐2) (Reichle et al., 2017) and Climate Forecast System Reanalysis (CFSR) (Dile & Srinivasan, 2014); and satellite data (Tropical Rainfall Measuring Mission, TRMM) (Huffman et al., 2007). These data sets have relatively good spatial (Figure S1 in the supporting information) and temporal coverage (Table 2). The original spatial resolution of ERA‐Interim is 0.75° × 0.75°; we use the resampled data with a 0.125° × 0.125° resolution.
| Data set | Coverage | Category |
Spatial (temporal) resolution |
Time Span | References |
|---|---|---|---|---|---|
| APHRODITE (V1101) | Asia | Interpolation | 1. (daily) | 1961–2007 | Yatagai et al. (2012) |
| CRU | Global | Interpolation | 0.5° (monthly) | 1961–2010 | Harris et al. (2014) |
| CFSR | Global | Reanalysis | 0.3125° (daily) | 1979–2011 | Dile and Srinivasan (2014) |
| ERA‐Interim | Global | Reanalysis | 0.125° (daily) | 1979–2011 | Dee et al. (2011) |
| MERRA‐2 | Global | Reanalysis | 0.625° × 0.5° (hourly) | 1980–2011 | Reichle et al. (2017) |
| TRMM (3B43) | Global | Satellite | 0.25° (daily) | 1998–2011 | Huffman et al. (2007) |
- Note. Time span shows the evaluation period used in this study.
The HydroSHEDS void‐filled digital elevation model (approximately 90 m × 90 m resolution) was used in this study (Lehner et al., 2008) (Figure 1a). A land use/land cover (LULC) data set was created using Landsat TM/ETM+ satellite imagery by unsupervised classification followed by classification based on the interpretation of Google Earth imagery (Figure 1b; overall accuracy 89%). A soil map (1:1,000,000) was obtained from the Institute of Soil Science, Chinese Academy of Science (CAS) (Shi et al., 2004) (Figure 1c). Soil texture parameters were derived from the Soil Map of China (National Soil Survey Office, 1995) and field sampling in September 2014 (Figure 1a). Laboratory analysis was carried out by employing a Laser Particle Size Analyzer (Malvern Mastersizer 3000). A lithology data set was derived from the 1:2,500,000 scale geological map of China (Figure 1d).
3 Methods
3.1 Gridded Data Sets Comparison and Correction
Gridded products need to be evaluated before modeling. As precipitation is the most important uncertainty source in mountainous regions (Hu et al., 2017), here, we evaluate the accuracy of six kinds of gridded precipitation products in and around the southern Tianshan where the Kaidu Basin is located. Precipitation extracted from gridded precipitation products was compiled at annual and monthly scales and compared directly with the nearest neighbor weather stations (Table 1). Correlation coefficient (R), standard deviation (SD), and root‐mean‐square (RMS) difference were presented in the Taylor diagrams (Taylor, 2001) in order to assess the feasibility of multiple gridded precipitation data in the Tianshan Mountains.
APHRODITE precipitation data were chosen eventually in terms of the best performance of seasonal distributions and time series dynamics at annual and seasonal scales (see section 4.1). However, APHRODITE has been reported to underestimate precipitation in mountain regions (Krysanova et al., 2015; Shea et al., 2015). In the Kaidu Basin, the mean annual precipitation from AHPRODITE is 277 mm (1972–2007) while the recorded annual discharge at the Dashankou gauging station is 188 mm (1972–2007) and the ActET is approximately 198 mm/yr (2001–2013) based on remote sensing estimation (Liu et al., 2017), which implies that APHRODITE underestimates precipitation in the Kaidu Basin based on the water balance. Precipitation shows spatial variation and a substantial altitudinal gradient caused by orographic effects (Chen, 2014). As the elevation of gridded APHRODITE differs from the measurement stations, we took the altitudinal gradient into account and statistically adjusted the precipitation in the Kaidu Basin.
ERA‐Interim temperature data offer the potential to fill data gaps in the mountain regions after elevation correction (Gao et al., 2012). Here ERA‐Interim temperature (maximum and minimum temperature) data were bias‐corrected using 1 year HOBO logger temperature data set (September 2014 to August 2015) (Figure 1a and Table 1) as suggested in previous studies (Gao et al., 2012; Shea et al., 2015). Daily corrected gridded temperature data can be calculated as:
| January | February | March | April | May | June | July | August | September | October | November | December | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
Lapse rate (°C/km) |
3.7 | 3.7 | 3.7 | 6.5 | 6.1 | 6.3 | 5.8 | 5.4 | 5.0 | 3.7 | 3.7 | 3.7 |
- Note. Temperature lapse rate from November to March were kept the same as October due to winter temperature inversion.
Relative humidity, solar radiation, and wind speed data sets were furthermore extracted from ERA‐Interim. They showed no strong relationship with station data, which can be attributed to the difference between grid box and point measurement. However, to share the same model physical mechanics with temperature, these data sets were kept unchanged.
3.2 Hydrological Model
The distributed and process‐based J2000 model (Krause, 2002), which is built upon the Jena Adaptable Modeling System (JAMS) (Kralisch et al., 2007; Kralisch & Krause, 2006), includes flexible components and modules for representing hydrological processes at the basin scale. It was successfully applied to simulate hydrological processes in mountainous regions (Biskop et al., 2016; Nepal et al., 2014), and therefore fits this study's purpose. The spatial heterogeneity of data and processes in the basin is represented by means of Hydrological Response Units (HRUs) (Flügel, 1995). HRUs were delineated by overlaying DEM derived information (elevation, slope angle, and slope aspect), lithology, land cover information, and soil data sets. All data sets were resampled to the spatial resolution of the DEM (90 m × 90 m), resulting in a total of 19,206 HRUs (Figure S3 in the supporting information). A short description of different modules is given below. Detailed information about the J2000 can be found in references on model design (Krause, 2002; Nepal, 2012).
Climate inputs are mapped to HRUs by means of a spatial regionalization approach based on inverse distance weighting and elevation regression. Precipitation is distributed into rain, snow, and rain‐snow mixtures by means of daily mean temperature and calibration parameters (Trs and Trans; Table 4). Maximum interception storage by vegetation coverage is modeled based on the Leaf Area Index (LAI) and its seasonal variability (Dickinson, 1984). From climate inputs, potential evapotranspiration is calculated according to the Penman‐Monteith method (Allen et al., 1998).
| Modules | Parameter acronym | Description (Units) | Range | Calibrated value |
|---|---|---|---|---|
| Initializing | ACAdaptation | Multiplier for air capacity (−) | 0.5–2 | 1.26 |
| FCAdaptation | Multiplier for field capacity (−) | 0.5–2 | 1.79 | |
| Precipitation distribution | Trs | Temperature threshold for snow and rain (°C) | −1 to +1 | −0.92 |
| Trans | Temperature range for mixed rain and snow (°C) | 0–2 | 0.48 | |
| Interception module | a_rain | Interception storage factor for rain (mm) | 0–5 | 0.60 |
| a_snow | Interception storage factor for snow (mm) | 0–5 | 1.09 | |
| Snow module | snowCritDens | Critical snow density (g/cm3) | 0.1–1 | 0.29 |
| baseTemp | Threshold temperature for snowmelt (°C) | −1 to 1 | −0.7 | |
| t_factor | Melt factor by sensible heat (mm/K) | 1–5 | 4.77 | |
| r_factor | Melt factor by liquid precipitation (−) | 1–5 | 2.09 | |
| g_factor | Melt factor by soil heat flow (mm) | 1–10 | 9.48 | |
| ccf | cold content factor (−) | 0.001–0.01 | 0.0014 | |
| Glacier module | meltfactor | Melt factor for ice melt (mm/K) | 0.2–5 | 0.32 |
| alphaIce | Radiation melt factor for ice (mm/(K * M)) | 0.1–1 | 0.1 | |
| ddfIce | Day degree factor for ice melt (mm/K) | 0.1–2 | 1.85 | |
| kIce | Routing coefficient for ice melt (−) | 1–8 | 2.1 | |
| kSnow | Routing coefficient for snowmelt (−) | 1–8 | 3.58 | |
| kRain | Routing coefficient for rain runoff (−) | 1–8 | 3.63 | |
| debrisFactor | Debris factor for ice melt (−) | 1–5 | 2.52 | |
| Tbase | Threshold temperature for melt (°C) | −2 to 2 | −0.25 | |
| Soil module | soilMaxDPS | Maximum depression storage (mm) | 0.3–2 | 1.73 |
| soilPolRed | polynomial reduction coefficient for actual evapotranspiration (−) | 1–5 | 2.96 | |
| soilMaxInfSummer | Maximum infiltration in summer (mm) | 60–150 | 82.10 | |
| soilMaxInfWinter | Maximum infiltration in winter (mm) | 60–150 | 65.15 | |
| soilMaxInfSnow | Maximum infiltration in snow cover areas (mm) | 40–120 | 43.19 | |
| soilImpGT80 | Infiltration for areas greater than 80% sealing (−) | 0.1–1 | 0.34 | |
| soilImpLT80 | Infiltration for areas lesser than 80% sealing (−) | 0.1–1 | 0.30 | |
| SoilDistMPSLPS | MPS–LPS distribution coefficient (−) | 0.2–2 | 0.64 | |
| SoilDiffMPSLPS | MPS–LPS diffusion coefficient (−) | 0.1–1 | 0.13 | |
| soilOutLPS | Outflow coefficient for LPS (−) | 0.1–1 | 0.23 | |
| soilLatVertLPS | Calibration coefficient for the distribution of interflow and percolation water (−) | 0.2–2 | 0.87 | |
| soilMaxPerc | Maximum percolation rate (mm) | 3–20 | 8.34 | |
| soilConcRD1 | Recession coefficient for overland flow (−) | 1–3 | 1.25 | |
| soilConcRD2 | Recession coefficient for interflow (−) | 1–4 | 2.8 | |
| Groundwater module | gwRG1RG2dist | RG1–RG2 distribution coefficient (−) | 0.5–3 | 1.39 |
| gwRG1Fact | Adaptation for RG1 flow (−) | 0.5–3 | 1.80 | |
| gwRG2Fact | Adaptation for RG2 flow (−) | 0.5–5 | 2.19 | |
| Reach routing | flowRouteTA | Run time of the outflow route (−) | 2–18 | 14.43 |
- Note. Parameters included in the sensitivity analysis are shown in boldface.
Snowmelt is simulated based on the approach suggested by Knauf (1980). As water from rain or snow can be stored in the snowpack, runoff from snowmelt occurs only when the storage capacity of the snowpack is exceeded. Thus, snow accumulation, compaction, and melt phases are included in the J2000 model. Snowmelt calculation considers two water fractions and snow densities: dry snow with an initial snow density and dry snow plus the stored liquid water with a modified snow density. This represents the ability of the snowpack to store liquid water without producing snowmelt runoff (Bertle, 1966; Krause, 2002). Even though snow sublimation affects evapotranspiration (Li et al., 2017), it was neglected in the model due to limited input data and should be regarded as a limitation.
A glacier simulation component was integrated into the model by using a degree‐day‐factor method (Hock, 1999), yet including more state variables (slope aspect, slope angle, and debris cover; Nepal, 2012). Snow processes on glaciers were modeled by the snowmelt model described above, while glacier melt only occurs once the surface snow has disappeared and the air temperature is higher than the glacier melt threshold. The glacierized HRUs are treated separately in the model, allowing glacier meltwater to create surface runoff directly. Glacierized HRUs with a slope angle below 30° are regarded as debris‐coverd glaciers, allowing to account for their modified energy balance in the model. The glacier extent is represented as being constant over time in the J2000 model (Nepal, 2012; Nepal et al., 2014).
Streamflow is calculated from four different runoff components which are simulated in the J2000, i.e., surface runoff (RD1), interflow from soil zone (RD2), interflow from the upper part of the aquifer (RG1), and base flow (RG2; Figure 2; Krause, 2002). Using this approach, both lateral and vertical fluxes are represented in the model. RD1 represents runoff from depression storage (DPS) which in turn is filled by infiltration and saturation excess water. Additionally, RD1 is fed by glacier runoff. Infiltrated water is distributed into two soil storages: the middle pore storage (MPS; pore diameter of 0.2–50 µm) and large pore storage (LPS; pore diameter >50 µm). Water stored in MPS is depleted by evapotranspiration while LPS is emptied by gravity. RD2 represents lateral flow from LPS in the soil layer, which reacts slower than RD1. Water from LPS can further percolate into the groundwater zone, representing the vertical water flux from the soil. The groundwater zone feds two additional runoff components, namely the faster (RG1) and the slower (RG2) groundwater fluxes. Processes related to permafrost are not explicitly represented in the model.
General schematic representation of the J2000 model based on Krause (2001).
The J2000 model takes reach routing into account based on the kinematic wave approach (Miller, 1984). The rate and velocity of flow in each stream segment was estimated by the Manning‐Strickler equation (Krause, 2001). A routing coefficient (TA) which is subject of model calibration influences the velocity of the runoff waves in the channel until it reaches the catchment outlet.
3.3 Model Calibration, Validation, and Uncertainty Analysis
A multiobjective genetic algorithm (Non‐dominated Sorting Genetic Algorithm‐II, NSGA‐II) was used to optimize model parameters (Deb et al., 2002). Three thousand simulations were performed with user‐defined parameter ranges (Table 4). The considered performance criteria were the Nash‐Sutcliffe efficiency (NSE; Nash & Sutcliffe, 1970), percent bias (PBIAS; Gupta et al., 1999), logarithmic Nash‐Sutcliffe efficiency (LNS; Krause et al., 2005), and coefficient of determination (r2). The time series of simulated and observed streamflow were split into subperiods based on reservoir operation: 1979–1981 for model initialization, 1982–1986 for calibration, and 1987–1991 (before reservoir construction) for validation. Considering the impacts of reservoirs since 1992 (Figure S4 in the supporting information), we do not expect the simulation to capture the hydrograph well after reservoir construction; simulation results for the 1992–2007 period are therefore shown only for information.
As different parameter sets may lead to equally accurate predictions, the prediction uncertainty was evaluated using the Generalized Likelihood Uncertainty Estimation (GLUE) procedure (Beven & Binley, 1992). The general idea behind GLUE is to run the model with random parameter sets, selecting behavioral Monte‐Carlo simulations while ruling out nonbehavioral ones in further analyses (Beven & Binley, 1992; Beven & Freer, 2001). It should be noted that simulations that are acceptable based on a goodness of fit criterion are referred to as behavioral models, although the acceptance threshold is subjectively determined. Each set of parameters from the behavioral simulations are assigned likelihood values, while unrealistic simulations are assigned zero. Finally, a measure of uncertainty of predictions is obtained based on the behavioral simulations. OPTAS (Fischer, 2013) with 7,734 Monte‐Carlo simulations was applied based on the parameter ranges (Table 4). To avoid the influence of reservoirs, the GLUE method was applied in the period of 1982–1991. In this study, simulations with LNS > 0.7 were considered as behavioral simulations. Based on this criterion, 80 parameter combinations were chosen as the behavioral ensemble. The 5th and 95th percentile was used to characterize the parameter uncertainties of the total streamflow and runoff components (RD1, RD2, RG1, and RG2). As meltwater is an important water source in the Kaidu Basin, uncertainties of snowmelt and glacier melt were also assessed separately using the GLUE method.
A Regional Sensitivity Analysis (RSA) was furthermore performed (Hornberger & Spear, 1981) to assess the parameter sensitivity of simulated streamflow. Monte‐Carlo simulations were implemented in the RSA analysis based on parameter ranges (Table 4). The RSA method separates various model simulations into behavioral (good) and nonbehavioral (bad) groups based on user‐defined evaluation criteria. The cumulative distributions of a single parameter associated with many model simulations are therefore indicators of parameter sensitivity. Large discrepancies in cumulative frequency distributions indicate a higher sensitivity (Fischer et al., 2012; Nepal et al., 2014). Sensitive parameters (Table 4) for the model output were singled out and discussed below (section 4.3).
4 Results
4.1 Evaluation of the Precipitation Products
4.1.1 Spatial Distribution of Precipitation Products
The spatial distribution of precipitation in the different gridded products is shown in Figures 3 and 4. Although the averaging periods from different gridded products are different (but overlapping), we believe that this should not impact the spatial distribution of precipitation fundamentally. Most of the gridded products can generally capture the spatial distribution of precipitation with higher values in the Tianshan Mountains in the northern and western part of the Kaidu Basin (Figure 3). CFSR, ERA‐Interim, MERRA‐2, and TRMM products generally overestimate precipitation in the western and northern parts of the Kaidu Basin where precipitation is approximately 400–800 mm per year (Chen, 2014). Particularly, annual CFSR, MERRA‐2, and TRMM have some extreme values (from 1,000 mm to more than 3,000 mm) which are unrealistic in this semiarid region (Figure 3). APHRODITE and CRU data have lower precipitation compared with the other precipitation data sets, yet CRU data mismatch the spatial pattern of precipitation at high elevations in general.
Spatial patterns of mean annual precipitation (mm) from the APHRODITE (1961–2007), CFSR (1979–2011), CRU (1961–2010), ERA‐Interim (1979–2011), MERRA‐2 (1980–2011), and TRMM (1998–2011) products over the Tianshan Mountains (81–88°E, 41–45°N).
Same as Figure 3 but for seasonal mean precipitation.
Overall, precipitation products can capture known seasonal precipitation patterns with a higher precipitation in summer and lower precipitation in winter (Chen, 2014) (Figure 4). Consistent with annual distribution, seasonal CFSR, ERA‐Interim, MERRA‐2, and TRMM overestimate summer precipitation which is even higher than the indicated sum of the entire year as mentioned above. To our knowledge, CFSR, ERA‐Interim, MERRA‐2, and TRMM data mostly overestimate spring precipitation, especially for CFSR (Figure 4). APHRODITE and CRU data have relatively reliable performance at seasonal scales for this semiarid mountain region. However, CRU showed less summer precipitation in the mountain chain.
4.1.2 Grid‐Based Comparison of Gridded Precipitation Products
Time series of precipitation at six weather stations were compared directly to the corresponding gridded data from each gridded product at annual and seasonal scales (Figures 5 and 6). Annual TRMM data overestimated precipitation at all stations and had an abruptly decreasing slope after the end of 1990s. The nonsystematic overestimation or underestimation errors and extremes at different stations indicate an insufficient quality of ERA‐Interim and CFSR data in this study region. MERRA‐2 data showed better annual consistency at the lower elevation stations than the mountain region (e.g., Bayinbuluke station). APHRODITE and CRU generally showed similar consistency with observation at most stations (Figure 5). Seasonally, most data sets captured the precipitation variability better than ERA‐Interim and CFSR, which had the precipitation maximum in June while the largest precipitation occurs in July normally (Figure 6). MERRA‐2 does not perform well, as it overestimates seasonal precipitation at the Bayinbuluke and Luntai stations and exhibits implausible seasonality at the Yanqi and Kuerle stations. TRMM revealed the highest positive bias in summer and exhibited unrealistic behavior in spring. CRU underestimates the summer precipitation at all the stations. APHRODITE is generally close to the observational data although it underestimates summer precipitation at the mountainous Bayinbuluke station.
Time series of precipitation at six weather stations (1961–2011) compared with gridded precipitation products from APHRODITE (1961–2011), CFSR (1979–2011), CRU (1961–2010), ERA‐Interim (1979–2011), MERRA‐2 (1980–2011), and TRMM (1998–2011).
Same as Fig 5 but for monthly mean precipitation.
Taylor diagrams were applied to evaluate all the comparable data sets except for TRMM, which was excluded from the comparison due to the short overlapping time period (Figure 7). ERA‐Interim and CFSR are unreliable at all the stations, having a higher SD and RMS and relatively weak correlations with precipitation at weather stations. MERRA‐2 generally had higher SD and RMS than CRU at all the stations, and had the highest SD and RMS overall at the Bayinbuluke mountainous station. APHRODITE performed best in terms of the higher correlation, lower SD and RMS at all the stations at annual scale.
Taylor diagrams for displaying correlation coefficient, SD and RMS of mean annual precipitation from observational stations and different gridded products based on the overlapping period 1979–2007. The azimuthal position gives correlation coefficient. The blue radial coordinates and the green concentric semicircles indicate SD and RMS values, respectively.
Seasonally, most gridded products had a dissatisfactory performance (Figures S5–S8 in supporting information). CFSR and ERA‐Interim had higher SD and RMS than the other data sets, and the correlations with the station data were below 0.5 in nearly all the seasons, except for the Bayinbuluke station. CRU and MERRA‐2 were inconsistent in most seasons and at various stations. At the mountainous Bayinbuluke station, MERRA‐2 had the highest SD and RMS in all the seasons. The highest correlation coefficient and smallest RMS were found in APHRODITE in all the seasons. Overall, APHRODITE achieved a good performance at annual and monthly scales and performed substantially better than the other available gridded precipitation products.
4.2 Temperature and Precipitation Correction
At all HOBO stations, the monthly mean temperature from ERA‐Interim was greater than the observed data. The gradients of bias were higher in winter time (November to March) (Figure S9 in the supporting information). We corrected ERA‐Interim temperature based on monthly mean bias from these nine HOBO stations. The bias‐corrected daily temperatures had strong correlations with the HOBO station (Figure 8), with root mean squared errors (RMSE) of 2.2 to 5.1°C mean absolute errors (MAE) of 1.8 to 3.7°C.
Relationships between mean daily temperatures observed at HOBO sites and bias‐corrected ERA‐Interim temperature (September 2014 to August 2015).
The adjusted precipitation exhibited a strong increase (nearly 50% compared to Bayinbuluke station) based on the elevation difference. The adjusted mean annual precipitation in the Kaidu Basin was 425 mm (discussed in section 4.6, Figure 14a), which is generally consistent with precipitation ranges of 200–500 mm mentioned by previous study (Fu et al., 2013). Spatially, regionalized precipitation is consistent with the general precipitation distribution in the Tianshan Mountains where the western and northern Tianshan receive more precipitation (400–800 mm) (Chen, 2014). Although substantial uncertainties remain, our experience so far shows that it is necessary to take the precipitation gradient into consideration, and the adjusted precipitation can fulfill the model requirement and represent the orographic effect.
4.3 Model Performance
Driven by the gridded precipitation and temperature data sets, the J2000 model could generally represent the regional hydrological dynamics, with NSE values of 0.69 and 0.61 and r2 values of 0.69 and 0.73 for the calibration and validation periods, respectively (Figures 9a and 9b). Simulated and observed monthly streamflow during calibration and validation period indicate a reasonably good fit as well (Figure S10 in the supporting information). For the postreservoir period, the NSE and R2 are 0.56 and 0.65, respectively (Figure 9c). It is worth noting that the model is not capturing reservoir effects, thus, we do not expect the model simulation to capture the hydrograph as well. The postreservoir period is therefore only shown for information.
Simulated and observed daily streamflow during the (a) calibration period of 1982–1986, (b) validation period of 1987–1991, and (c) postreservoir period of 1992–2007 in the Kaidu Basin. Performance measures were calculated from daily data.
Although the rising and recession limbs are generally captured by the model, the spring snowmelt in the calibration period was underestimated (Figure 9a). However, the simulated base flow is relatively reliable in the basin based on the LNS, which is more closely related to the low flow. The LNS values are 0.79 and 0.84 for the calibration and validation periods (Figures 9a and 9b). Additionally, some overprediction peaks still can be seen in the postreservoir periods when the model is not intended to simulate regulated flow due to reservoirs.
4.4 Uncertainty and Sensitivity Analysis
Based on the GLUE uncertainty analysis, the observed streamflow is generally within the 5th to 95th percentile of uncertainty ranges (1982–1991; Figure 10a). However, the uncertainty ranges are not constant in the simulated years; they are larger in the summer (high flow) periods when both precipitation and snow/glacier melt play an important role while they are relatively small in the winter (low flow) periods (Figure 10b). Furthermore, the streamflow in April and October is not well represented by the simulations as it falls outside of the uncertainty ranges (Figure 10b).
(a) Uncertainty band and observed streamflow during the period of 1982–1991 based on the GLUE method. (b) Boxplot of monthly uncertainty band of simulated streamflow (1982–1991). Boxplots represent extreme values, lower and upper quartiles and median value of a variable (similarly hereinafter).
Figure 11 shows the average dynamics in the centerline of the box of the behavioral models for snowmelt, glacier melt, and the runoff components RD1, RD2, RG1, and RG2. The range of the boxes and whiskers can be interpreted by GLUE‐based uncertainty. The snowmelt shows two peak dynamics with maximum values in May and September. The uncertainty in May (±16%) is lower than in April and June (±38% and ±35%, respectively) while the uncertainties from August to October are ±56%, ±31%, and ±30%, respectively. The uncertainty in simulated glacier melt is more obvious in July and August (±61% and ±50%, respectively). The patterns of RD1, RD2, and RG1 are generally similar, with the absolute uncertainty range higher in summer months. The annual average relative uncertainties are ±34%, ±53%, and ±43%, respectively. RG2 has a more stable uncertainty (±21%) throughout the year.
Boxplots of the effect of parameter uncertainties on different simulated runoff components (1982–1991). Note: the plot of snowmelt is different with other figures.
Model performance was particularly sensitive to four parameters (flowRouteTA, alphaIce, gwRG1RG2dist, ccf; Figure 12) based on RSA analysis. The model performance is moderately sensitive to additional parameters of the groundwater module (gwRG2Fact), the glacier module (meltFactorIce, debrisFactor, and tbase) and the snowmelt module (g_factor, t_factor). The remaining parameters are less sensitive based on the LNS evaluation criterion (Figure 12).
Parameter sensitivity based on the LNS evaluation criterion.
4.5 Temporal Variation of the Simulated Water Balance and Runoff Components
Annual and monthly distributions of the simulated water balance are shown in Table 5 and Figure 13a. As major inputs for the water balance, precipitation and glacier melt account for 425 mm and 9 mm, respectively. As for the output from the hydrological system, ActET, and simulated streamflow account for 233 mm and 199 mm, respectively. Annual storage changes according to the model sum up to 2 mm. The summer season dominates the major hydrological processes in the Kaidu Basin, with peak values of precipitation, temperature, ActET, and runoff in summer (Table 5).
| Precipitation | Glacier melt | ActET | Simulated runoff | Storage change | |
|---|---|---|---|---|---|
| Jan | 8.1 | 0.0 | 3.7 | 8.1 | −3.6 |
| Feb | 8.9 | 0.0 | 4.3 | 7.1 | −2.6 |
| Mar | 11.7 | 0.0 | 11.4 | 7.7 | −7.3 |
| Apr | 20.2 | 0.0 | 24.1 | 12.6 | −16.5 |
| May | 43.9 | 0.1 | 30.21 | 19.0 | −5.2 |
| Jun | 88.4 | 1.4 | 35.8 | 26.6 | 27.4 |
| Jul | 92.3 | 3.1 | 41.8 | 32.0 | 21.6 |
| Aug | 74.3 | 3.0 | 35.6 | 32.8 | 8.8 |
| Sep | 37.3 | 1.1 | 22.5 | 22.1 | −6.2 |
| Oct | 14.9 | 0.2 | 12.4 | 13.3 | −10.6 |
| Nov | 13.9 | 0.0 | 7.0 | 9.4 | −2.5 |
| Dec | 11.3 | 0.0 | 4.0 | 8.6 | −1.2 |
| Annual | 424.9 | 8.9 | 232.6 | 199.2 | 2.0 |
- Note. Storage change represents changes in channel, soil layer, snow cover, groundwater, and surface storages. Values are in mm.
(a) Simulated water balance, (b) streamflow, snowmelt and glacier melt, and (c) distributions runoff components in the Kaidu Basin (1982–2007).
As important water sources in the glacierized basin, snowmelt, and glacier melt were extracted and analyzed individually (Figure 13b). Snowmelt mainly takes place in April (26 mm) and May (17 mm), while glacier melt is most evident in July and August (3 mm each). Snowmelt in September (8 mm) and October (6 mm) are also remarkable. Overall, snowmelt and glacier melt contribute 33% (66 mm) and 5% (9 mm) to the simulated annual streamflow, respectively.
Concerning the monthly distribution of different runoff components, RG2 (91 mm) and RD1 (62 mm) are the most important contributors, which account for 46% and 31% of the simulated annual total streamflow, respectively (Figure 13a). RD2 (13 mm) and RG1 (32 mm) account for 7% and 16% of the simulated annual total streamflow, respectively. Additionally, RD1, RD2, and RG1 showed similar seasonal patterns with peaks in summer and lows in winter, while RG2 was relatively stable throughout the whole year (Figure 13c).
4.6 Spatial Characteristics of Simulated Water Balance and Runoff Components
To explore the spatial distribution of water balance and runoff components, we aggregated these components to mean annual values (Figure 14). Mean annual precipitation varies regionally from 223 to 560 mm in the basin, with more humid areas located in the western and northern parts while the drier areas are in the southeastern basin (Figure 14a). The simulated ActET is partly impacted by elevation, with a lower ActET in the mountain regions. The highest ActET was concentrated in the lower reaches of the river and in the middle of the Kaidu Basin where wetlands are located (Figure 14b). The ActET rates are lower in the southeastern basin than in the higher elevated central part due to the limited water supply. The spots of lower ActET rates in the northeastern plane are caused by the barren land which leads to relatively lower ActET values. The annual streamflow is mainly generated in the mountain regions in the middle, western, and northern part of the Kaidu Basin (Figure 14c). The snowmelt distribution shows higher values in the western mountain regions (Figure 14d). Apart from the greater water availability due to higher precipitation rates, RD1 and RD2 are mainly generated in the mountain regions, which is associated with higher slopes and less developed soils (Figures 14e and 14f). Barren land increases RD1 because of lower infiltration rates while gentle slopes and well‐developed soils are associated with less RD2 in the lowlands. As the distribution of RG1 and RG2 depends on the HRUs' slope gradients, less RG1 is generated in the plain areas (Figure 14g). The distribution of modeled RG2 mainly depends on the total water balance. The changes of RG2 can be regarded as compensation with other runoff components, which can be observed in barren land where RD1 is greatly generated while RG2 is less distributed (Figures 14e and 14h).
Spatial distribution of simulated mean annual (a) precipitation, (b) ActET, (c) streamflow, (d) snowmelt, (e) RD1, (f) RD2, (g) RG1, and (h) RG2 in the Kaidu Basin (1982–2007).
5 Discussion
5.1 Comparison and Correction of Gridded Meteorological Data Products
No consistent error pattern could be identified in the comparison of several gridded precipitation products and weather stations in our study (Figures 3-7). As different gridded data sets utilized different assimilation methodologies, we believe that the nonsystematic errors can be due to limitations of the data assimilation processes and lack of representative weather stations in this mountainous region. However, gridded data sets need to be evaluated and used carefully, and the reasons for the spatial and temporal errors in different gridded data products need further study.
APHRODITE performed better in terms of an acceptable SD, RSM, and R, which is related to the fact that APHRODITE was produced by interpolating rain gauge stations (Yatagai et al., 2012). However, APHRODITE underestimates precipitation in mountainous regions. Precise estimates of the amount of precipitation in the Kaidu Basin are not yet available due to scarcity of observational data and complex terrain. We have considered a spatial downscaling methodology, yet it will lead to unrealistic results in this data‐scarce mountain region. The precipitation factor method has been proposed as an acceptable option in data‐scarce mountain regions (Biskop et al., 2016; Immerzeel et al., 2015). APHRODITE was therefore simply corrected by a basin‐wide factor (1.5) based on the annual precipitation gradient and elevation difference between the grid box and weather station. We adopted this approach because it may be more accurate in complex terrain where precipitation generally increases with elevation and it also provided a clear improvement in simulating regional streamflow in the Kaidu Basin. The adjustment of precipitation provides a plausible estimation of annual precipitation, resulting in the simulated ActET being in much closer agreement with a previous study by Liu et al. (2017). These results support the chosen precipitation correction and modeling approach. However, uncertainty remains in the precipitation input.
A constant lapse rate is often a poor description of the spatial or temporal temperature structure (Immerzeel et al., 2014; Lundquist & Cayan, 2007). Thus, monthly temperature lapse rates were used for temperature correction and hydrological modeling in this study. Similar bias correction methods have frequently been used in the literature (Gao et al., 2012; Liu et al., 2011; Shea et al., 2015). The largest bias in winter temperature during the correction processes might be interpreted as a temperature inversion (Shen et al., 2016) which can be captured by local monitoring stations but not the ERA‐Interim temperature data.
5.2 Remarks on Calibration, Uncertainty, and Sensitivity Analysis
The performance of hydrological models in glacierized catchments is strongly constrained by various uncertainties: quality of the input data, model structure, model parameters, and the limited knowledge of hydrological regimes (Chen et al., 2016b; Montanari et al., 2009; Nepal et al., 2014; Shea et al., 2015; Ragettli et al., 2013). Existing studies utilized different model approaches for modeling snowmelt in different basins, which makes the parameters nontransferable (Dou et al., 2011; Zhang et al., 2007, 2015). Some of the model parameters were derived from modeling studies in other mountain regions (Biskop et al., 2016; Nepal et al., 2014; Ragettli et al., 2015) and adapted to our regional setting. However, calibration parameters in the model are often related to each other (Figure 2 and Table 4). Their values therefore need to be chosen with care. Although we defined plausible parameter ranges based on the literature and our knowledge about the study region, different parameter sets might lead to the same prediction as suggested by Beven and Freer (2001). Indeed, there is not a single “best” parameter set, and we therefore conclude that hydrological regimes can best be represented by the behavioral parameter ensemble based on calibration and uncertainty analysis.
Although previous research indicated that model nonlinearity, parameter uncertainty, and model structure errors can be implicitly reflected in the GLUE method (Beven & Binley, 1992; Beven & Freer, 2001), the GLUE method has still been criticized for its subjective selection of behavioral models (Mantovan & Todini, 2006) and inconsistency with statistical theory (Montanari et al., 2009; Stedinger et al., 2008). However, subjectivity is unavoidable and expert knowledge is highly needed where data are scarce (Beven & Binley, 2014; Montanari et al., 2009). Given that epistemic errors make it hard to fit the assumptions of probabilistic models in real‐world applications, GLUE remains a widely used technique applicable in data‐scarce regions we therefore believe that the GLUE uncertainty assessment is acceptable in this study while its limitations should be acknowledged.
Snowmelt is too complex to be fully represented due to data scarcity and some unpredictable processes (e.g., snow drift), which may be the reason for the large range of snowmelt uncertainties (Figure 11) and underprediction of spring snowmelt in the calibration period (Figure 9a). In absolute terms, the great summer uncertainties of glacier melt can be explained by the dynamics of temperature and precipitation, which impact glacier melt the most (Figure 11). Runoff components RD1, RD2, and RG1 show a larger uncertainty range in summer than in winter time, which is due to the fact that surface and subsurface runoff react relatively quickly in summer to the changes of hydrometeorological conditions. In addition, the variability of precipitation and glacier melt in the summer season could strengthen these runoff uncertainties. Uncertainty of the base flow component RG2 was less pronounced, which is not surprising since base flow varies more slowly in the water cycle (Nepal et al., 2014).
The most sensitive parameters were identified based on RSA and provide starting points for further reducing model uncertainty in future studies (Table 4). The Kaidu Basin is a large basin in which large portions of the stream network have weak gradients. Streamflow is therefore greatly affected by the calibration parameter flowRouteTA, which refers to the velocity of the streamflow waves (Figure 12). Base flow plays an important role in sustaining streamflow, especially in the winter time. The gwRG2Fact and gwRG1RG2dist influence the runoff retention time and the distribution of water between RG1 and RG2, which is the reason for the sensitivity of model performance to these parameters. The water budgets of mountain areas also partly depend on meltwater from glaciers (Chen et al., 2016a). Streamflow variability is therefore also highly sensitive to alphaIce, meltfactor, debrisFactor, and tbase, which affect glacier melt calculation directly in J2000 (Nepal et al., 2014). The sensitivity to ccf, t_factor, and g_factor can be explained by the importance of snow cold content in snow accumulation and snowmelt (Krause, 2002). The soil module dominates the water transfer between surface and subsurface, which plays an important role in the simulated streamflow. The simulated streamflow is therefore also sensitive to parameters of the soil module (Figure 2 and Table 4).
5.3 Simulated Water Balance and Runoff Components
Although uncertainties remain and it is impossible to validate the volume of annual precipitation in the Kaidu Basin without additional measurement data, the amount of precipitation in this study (approximately 425 mm per year) is in reasonable agreement with other studies according to which precipitation amounts range from about 200 to 500 mm in the Kaidu Basin (Fu et al., 2013). Spatially, the distribution of precipitation broadly agrees with the previously known precipitation patterns in the Tianshan Mountains according to which precipitation is higher in the adjacent northern Tianshan and the Yili River valley (400–800 mm per year; Chen, 2014), and the western Tianshan (Chinese part) receives more precipitation than the eastern part (Chen, 2014; Xu et al., 2010). The simulated mean annual ActET (233 mm) is slightly higher than the satellite‐based estimation (2001–2013) of mean annual ActET (above 190 mm) in the Kaidu Basin (Liu et al., 2017). Although little data is available to validate ActET in both our study and the satellite‐based estimation (MODIS), we believe the estimation of regional ActET can be trusted owing to the good performance of the hydrological model and the closed water balance.
The large amount of simulated base flow (46% of the total streamflow) suggests that groundwater related processes play an important role in the Kaidu Basin. The simulated base flow volumes are in good agreement with results from other study (approximate 41%; Chen et al., 2009). The seasonal distribution of base flow is stable, which is reasonable since groundwater has a longer recession time (Figure 13c). According to our model results most of the high‐intensity rainfall drained as surface flow due to the mountainous topography; this may explain the large contribution (31%) of RD1. The smaller contribution of RD2 (7%) is possibly due to poorly developed soils on the steep slopes where most of the surface runoff is generated. The proportion of infiltrated water is less due to the small pore storage, which saturates quickly and leads to a major contribution to surface runoff.
5.4 Implications and Future Work
Climate change substantially influences hydrological regimes in basins with high runoff contributions from snow and glacier meltwater, which could lead to more pressure on water availability in the future (Sorg et al., 2012). The water cycle has already started to intensify and could become more unstable under a warming climate in the Tianshan Mountains (Shen & Chen, 2010; Shi et al., 2007). Additionally, irrigation agriculture consumes large amounts of water further downstream and relies heavily on stream discharge, which is generated primarily by mountain meltwater or summer rainfall (Shen et al., 2013). However, snowmelt runoff occurs earlier in a warming climate (Liu et al., 2011; Sun et al., 2015), which may reduce water availability in summer when irrigation demand is at its peak. The water balance and the distributions of runoff components in the glacierized Kaidu Basin were quantified. Therefore, we believe the results of this study explained a significant part of mountain hydrology and could be helpful for adopting better water resource management.
This study is of interest as hydrological modeling in this data‐scarce basin has not previously been described in detail. Integrating meteorological data from multiple sources sheds light on modeling mountain hydrology, which could be important for model applications and design in data‐scarce mountainous regions elsewhere. The existence of parametric sensitivity and uncertainty could contribute to improving model calibration and design in Central Asia in the future. Our results also clearly highlight the need for adequate and sustainable observing systems for modeling present and future water resources.
6 Conclusion
In this study, we evaluated the performance of different gridded meteorological data sets in the Tianshan Mountains in order to identify the most appropriate inputs for driving hydrological models. Additionally, we statistically bias‐corrected gridded climate data based on in situ data and analyzed the spatiotemporal patterns of water balance and the distribution of runoff components in the glacierized Kaidu Basin using the J2000 model. Moreover, we have identified parameter uncertainties and sensitivities that affect model performance, which need to be considered and further constrained in future research.
Overall, none of the available gridded precipitation products is perfect, but the interpolated APHRODITE data represented the annual and seasonal precipitation dynamics best, although it underestimates precipitation in the Tianshan Mountains. Driven by the corrected gridded climate data sets, the simulated daily streamflow is in good agreement with observed streamflow in the calibration and validation periods. Parameter uncertainty and sensitivity analyses were conducted, which are essential for modeling studies in data‐scarce mountain basins. Regional water balance and the distribution of runoff components were further quantified, and uncertainties in streamflow and runoff components were estimated, which can be used to reduce hydrological uncertainties in future work. We acknowledge that uncertainty remains due to limited data availability. However, this study provides insights into mountain hydrology, which could contribute to hydrological research and water resource management in the Tianshan mountain area and other mountain regions with limited measurement data. Further studies are required to address the limitations and uncertainties.
Acknowledgments
The research is supported by the Key Research Program of National Natural Science Foundation of China (D91425302) and the National Natural Science Foundation of China (41630859). Thanks to China Scholarship Council (CSC) for a PhD scholarship (201304910343). The meteorological data sets and field data can be obtained from the Kaidu River Basin Information System (KaiduRBIS): http://leutra.geogr.uni-jena.de/kaiduRBIS/metadata/start.php. Colleagues from the Key Laboratory of Agricultural Water Resources, Center for Agricultural Resources Research, CAS, are gratefully acknowledged for their support during field survey. We thank Jason Goetz and Miga Magenika Julian for their valuable suggestions. We appreciate the editor and anonymous reviewers for their constructive comments.
