Advertisement

Climate Dynamics

, Volume 50, Issue 7–8, pp 2829–2844 | Cite as

A hybrid model to assess the impact of climate variability on streamflow for an ungauged mountainous basin

  • Chong Wang
  • Jianhua Xu
  • Yaning Chen
  • Ling Bai
  • Zhongsheng Chen
Article

Abstract

To quantitatively assess the impact of climate variability on streamflow in an ungauged mountainous basin is a difficult and challenging work. In this study, a hybrid model combing downscaling method based on earth data products, back propagation artificial neural networks (BPANN) and weights connection method was developed to explore an approach for solving this problem. To validate the applicability of the hybrid model, the Kumarik River and Toshkan River, two headwaters of the Aksu River, were employed to assess the impact of climate variability on streamflow by using this hybrid model. The conclusion is that the hybrid model presented a good performance, and the quantitative assessment results for the two headwaters are: (1) the precipitation respectively increased by 48.5 and 41.0 mm in the Kumarik catchment and Toshkan catchment, and the average annual temperature both increased by 0.1 °C in the two catchments during each decade from 1980 to 2012; (2) with the warming and wetting climate, the streamflow respectively increased 1.5 × 108 and 3.3 × 108 m3 per decade in the Kumarik River and the Toshkan River; and (3) the contribution of the temperature and precipitation to the streamflow, which were 64.01 ± 7.34, 35.99 ± 7.34 and 47.72 ± 8.10, 52.26 ± 8.10%, respectively in the Kumarik catchment and Toshkan catchment. Our study introduced a feasible hybrid model for the assessment of the impact of climate variability on streamflow, which can be used in the ungauged mountainous basin of Northwest China.

Keywords

Hybrid model Assessment Climate variability Streamflow Ungauged mountainous basin Northwest China 

1 Introduction

Global warming is an indisputable fact (Allen et al. 2009). As a result, glaciers shrink worldwide, Northern Hemisphere spring snow cover decreases in extent, precipitation in North Hemisphere mid-latitude land areas increases, and extreme weather and climate events rises (Pachauri et al. 2014). The headwaters of inland rivers in Northwest China are mainly recharged by glaciers, snow meltwater and precipitation (Chen et al. 2006). So the climate variability alters hydrological process and affects the quantity and quality of water resources in inland river basins (Guo and Shen 2016; Ma et al. 2008). However, most of the headwaters of the inland rivers are located in high-altitude ungauged mountains. It’s difficult to quantitatively assess the climate variability and its impact on streamflow. Hence, it’s necessary to develop a hybrid model to solve this problem.

Many researches are interested in the impact of climate variability on streamflow in inland river basins (Yang et al. 2015). Inland rivers are sensitive against climate variability (Xu et al. 2009) Take the Tarim Basin, for instance. The annual temperature increased by almost 1 °C in the basin since 1955, and the annual precipitation showed a rising trend at a rate of 6–8 mm per decade (Chen et al. 2006). Because of the increasing mountainous temperature and precipitation, glacier and snow meltwater increased, and the streamflow in tributaries of the Tarim River augmented (Chen et al. 2013; Xu et al. 2013a, b, c, d). However, the lack of meteorological stations limited the research accuracy in mountainous area such as headwaters.

Previous studies proposed many methods to reveal climate characteristics in ungauged mountains (Neupane et al. 2015; Zhao et al. 2016). The reanalysis earth data provided spatial distribution of climate (Getirana et al. 2011), but the low resolution eliminated the climate heterogeneity at the scale of catchment. HHMM (nonhomogeneous hidden Markov model), SDSM (statistical downscaling model) and other downscaling models could reconstruct the regional climate, but researches showed that the precipitation simulation accuracy still need to be improved (Liu et al. 2011; Mahmood and Babel 2013). It was also difficult to accurately describe the nonlinear relationship between the climate and streamflow with multivariate linear regression method (Zhang and Zuo 2015). Hydrological model based on physical process could simulate the climate and streamflow, but the operation was complex and the uncertainty was increasing with the adding data and parameters (Duethmann et al. 2015). It’s difficult to quantitatively assess the impact of climate variability on streamflow in the ungauged mountainous basin with vast area, scarce meteorological and hydrological observation stations and insufficient data in Northwest China.

To quantitatively assess the impact of climate variability on streamflow for an ungauged mountainous basin, We conducted a hybrid model integrating a downscaling method based on earth data products, Mann–Kendall trend test, back propagation artificial neural networks (BPANN) and weights connection method, and selected the headwater basins of the Aksu River (i.e. the Kumarik River and Toshkan River) as the typical ungauged basins to validate the model.

2 Materials and methods

2.1 Description of study area

As one of the tributaries of Tarim River in Northwest China, the Aksu River provides about 70–80% of water for the mainstream (Chen et al. 2003; Duethmann et al. 2015). However, the water of the Aksu River is mainly from the headwaters originating from Tianshan Mountains.

As the main headwaters of the Aksu River, the Kumarik River and Toshkan River are located on the southern slope of the western Tianshan Mountains (Fig. 1), which are in the typical ungauged basins with vast area, scarce meteorological and hydrological observation stations and insufficient data. So, we selected the two headwater basins as the study target. The latitude is from 75.5°E to 80.3°E, and the longitude is from 40.2°N to 42.6°N. The altitude of the Kumarik catchment is from 1409 to 7077 m with an average of 3730 m. In the Toshkan catchment, the altitude is from 1931 to 5934 m with an average of 3730 m. The areas of the two catchments are respectively 12,983 and 18,331 km2. The areas of glaciers are 2583 and 887 km2 in the two catchments. The study area has a typical temperate continental climate. Based on observations from Aheqi and Chatyr Kul (1980–2012), the average temperature is 1.9 °C and the average monthly temperature is below 0 °C from November to March next year. The annual precipitation is about 250 mm and the monthly precipitation is more than 30 mm from May to September. Glacier, snow meltwater and precipitation constitute the main components for the water supply to the headwaters. Grazing activities in the catchments are small and the influence of human activities is negligible in local hydrological process.

Fig. 1

The study area

2.2 Data sources

The terrain data is from the SRTM (Shuttle Radar Topography Mission) V4.1 preprocessed by CIAT (International Center for Tropical Agriculture) (Reuter et al. 2007). USGS (United States Geological Survey)/NASA (National Aeronautics and Space Administration) provided the data (http://srtm.csi.cgiar.org). The spatial resolution of the raster data is 90 m × 90 m.

The glacier data is from the GLIMS (Global Land Ice Measurements from Space) (Raup et al. 2007) provided by NSIDC (National Snow and Ice Data Center) (http://glims.colorado.edu). The monitoring time was on June 11, 2013. The data format is polygon vector.

The reanalysis data was downscaled to simulate climate change in the ungauged basins. After the accuracy test on a series of reanalysis data of temperature, the MERRA-2 instM_2d_lfo_Nx V5.12.4 data (http://goldsmr4.gesdisc.eosdis.nasa.gov) from GMAO (Global Modeling and Assimilation Office) was selected. The data format is netCDF-4 with a spatial resolution of 0.5° × 0.625° and monthly temporal resolution. The data begins in January 1980 and end in December 2012.

After the accuracy test on a variety of reanalysis data of precipitation, the China precipitation grid data V2.0 (http://data.cma.cn) from NMIC (National Meteorological Information Center) was selected. The data format is ARCGIS standard format with a spatial resolution of 0.5° × 0.5° and monthly temporal resolution. The data begins in January 1980 and end in December 2012.

In order to verify the accuracy of the downscaling model, the observed temperature and precipitation data from Chinese ground meteorological database (http://data.cma.cn), NMIC (National Meteorological Information Center) was used. The meteorological stations are Aksu, Keping, Aheqi and Chatyr Kul. The format is ASCII with monthly temporal resolution. The data begins in January 1980 and end in December 2012.

For the purpose of verifying the effect of stream simulation, the observed streamflow data from the monitoring records of hydrological stations provided by Xinjiang Tarim River Basin Management Bureau was chosen. The hydrological stations of the Kumarik River and Toshkan River are Xiehela and Shaliguilanke. The format is ASCII with monthly temporal resolution. The data begins in 1980 and end in 2012.

2.3 Methods

We conducted a hybrid model to assess the impact of climate variability on streamflow for an ungauged mountainous basin (Fig. 2). The model consists of three sub models. The first sub model downscales the earth data products and verifies the accuracy. The second sub model analyzes the climate variability. The last sub model constructs a BPANN (back propagation based artificial neural network) based on downscaled climate data and connects the weights. The final output is the quantitatively assessment of climate variability impact on streamflow.

Fig. 2

The framework of the hybrid model to assess the impact of climate variability on streamflow

2.3.1 Downscaling based on earth data products

There is no national basic weather station within the study catchment. Downscaling the reanalysis data is a feasible way to reconstruct the historical climate variability (Xu et al. 2013a, b, c, d) or provide temperature and precipitation estimation based on forecast models (Georgakakos et al. 2014). Liston and Elder (2006) proposed an effective downscaling method using the empirical temperature lapse rates and precipitation gradients. Brown et al. (2014) used this downscaling method in estimating glacier and snow melt driven streamflow in Himalayas. However, based on data visualization, we found that the real temperature lapse rates in Northwest China were generally higher than the empirical coefficients, and the real precipitation gradients changed with the altitude rather than fixed values. Based on previous research, we developed an improved downscaling method based on earth data products in the first sub model. The first sub model contains simulating temperature lapse rates and precipitation gradients, downscaling the reanalysis climate data to a high resolution (90 m × 90 m) and calculating the monthly temperature and precipitation in catchment (Fig. 3).

Fig. 3

Downscaling sub model based on earth data products

To simulate the temperature lapse rates and precipitation gradients, the sub model modeled the relationships between the temperature t, precipitation p and the altitude h at first. Because of the monthly temporal resolution of the reanalysis data, we chose the monthly temperature and precipitation observed data from meteorological stations (Aksu, Keping, Aheqi and Chatyr Kul) from 1980 to 2012 to construct the model. The monthly data were preprocessed to eliminate seasonal fluctuations and improve the fitting precision as follows:
$${t_i}=\mathop \sum \limits_{j=1}^{12} {t_m}_{{ij}}/12,$$
(1)
$${p_i}=\mathop \sum \limits_{j=1}^{12} {p_m}_{{ij}}/12,$$
(2)
where \({t_m}_{{ij}}\) is the monthly temperature in the ith year and jth month, \({p_m}_{{ij}}\) is the monthly precipitation in the ith year and jth month.
According to the scatter plots and previous studies in the similar areas (Fu et al. 2013), the relationship between temperature t and altitude h was linear, and the relationship between precipitation p and altitude h was quadratic polynomial. So we established a linear function \(t=mh+n\) and a quadratic polynomial function \({~}p={a}{h^2}+bh+c\) to simulate the temperature and precipitation. The coefficients m, n, a, b, c were fitted with observed data. The coefficient of determination R 2 and F test was used to verify the fitting effect. R 2 was calculated as the following Eq. (3):
$${R^2}=1 - \frac{{\mathop \sum \nolimits_{i=1}^k {{\left( {{y_i} - {{\hat y}_i}} \right)}^2}}}{{\mathop \sum \nolimits_{i=1}^k {{\left( {{y_i} - {{\bar y}_i}} \right)}^2}}},$$
(3)
where, \({y_i}\), \({\hat y_i}\), \({\bar y_i}\) are respectively observed, simulated and average observed values. k is the number of the observed data.
Then the temperature lapse rate \(\frac{{{d}t}}{{{d}h}}\) and the precipitation gradient \(\frac{{{d}t}}{{{d}h}}\) were calculated by differential of the fitting functions:
$$\frac{{{d}t}}{{{d}h}}=m,$$
(4)
$$\frac{{{d}p}}{{{d}h}}=2ah+b.$$
(5)
The sub model downscaled the monthly reanalysis climate data by calculating the changing values of temperature and precipitation according to altitude differences. At first, the sub model calculated the altitude of the reanalysis data h r according to the altitude of SRTM h s by bilinear interpolation resampling. Then the sub model resampled the reanalysis temperature data t r , reanalysis precipitation data p r and h r to the grid size of h s by bilinear interpolation. The preprocessing facilitated the matrix operation among different data type. To improve resolution, we need to change each pixel value of reanalysis data from its altitude to the altitude of SRTM data. The value changed with the altitude according to the temperature lapse rate and precipitation gradient. So finally, the downscaled climate data t d and p d (90 m × 90 m) at the altitude h s were calculated by Eqs. (6) and  (7):
$${t_d}={t_r}+\left( {{h_s} - {h_r}} \right) \cdot \left( {\frac{{{d}t}}{{{d}{h_s}}}+\frac{{{d}t}}{{{d}{h_r}}}} \right)/2,$$
(6)
$${p_d}={p_r}+\left( {{h_s} - {h_r}} \right) \cdot \left( {\frac{{{d}p}}{{{d}{h_s}}}+\frac{{{d}p}}{{{d}{h_r}}}} \right)/2,$$
(7)
where \(\left( {{h_s} - {h_r}} \right)\) is the altitude differences, \(\left( {\frac{{{d}t}}{{{d}{h_s}}}+\frac{{{d}t}}{{{d}{h_r}}}} \right)/2\) is the average temperature lapse rates between altitude \({h_s}\) and \({h_r}\), \(\left( {\frac{{{d}p}}{{{d}{h_s}}}+\frac{{{d}p}}{{{d}{h_r}}}} \right)/2\) is the average precipitation gradients between altitude \({h_s}\) and \({h_r}\).
With the Eq. (4) and Eq. (5), the final downscaling equations were:
$${t_d}={t_r}+\left( {{h_s} - {h_r}} \right) \cdot m,$$
(8)
$${p_d}={p_r}+\left( {{h_s} - {h_r}} \right) \cdot \left[ {a\left( {{h_s}+{h_r}} \right)+b} \right].$$
(9)
The sub model verified the downscaled data by calculating the evaluation indexes including Slope, NSE (Nash–Sutcliffe efficiency coefficient), MAE (mean absolute difference) and RMSE (root-mean-square error) between the observed climate data and the downscaled data at the same location. In order to intuitively reflect the fitting effect of downscaled monthly temperature and precipitation on observed values, we drew the scatter plot, fitted the trend line and calculated the R 2 . The downscaled values at location of the meteorological stations were extracted by nearest neighbor method from t d and p d . The evaluation indexes were calculated as the following equations:
$$NSE=1 - \frac{{\mathop \sum \nolimits_{i=1}^k {{\left( {{y_i} - {{\hat y}_i}} \right)}^2}}}{{\mathop \sum \nolimits_{i=1}^k {{\left( {y - {{\bar y}_i}} \right)}^2}}},$$
(10)
$$MAE=\mathop \sum \limits_{i=1}^k \left| {{y_i} - {{\hat y}_i}} \right|/k,$$
(11)
$$RMSE=\sqrt {\mathop \sum \limits_{i=1}^k {{\left( {{y_i} - {{\hat y}_i}} \right)}^2}/k}$$
(12)
where, \({y_i}\), \({\hat y_i}\), \({\bar y_i}\) are respectively the observed, downscaled and average observed data and k is the number of observed data.

2.3.2 Analysis of climate variability

To make a statistical description of the climate variability and test the trend in each catchment, the second sub model calculated the monthly temperature and precipitation in catchments by arithmetic average method. Then the annual average temperature and annual total precipitation in catchments were calculated. The statistical description consisted of Mean, SD (standard deviation), CV (the coefficient of variation) and slope (linear trend). The slope was calculated by fitting the linear function between the climate factors and the year based on least squares method.

To test the significance of the slope, the sub model employed the Mann–Kendall trend test (Kendall 1948; Mann 1945). The Mann–Kendall trend test was widely applied in meteorological and hydrological researches (Bai et al. 2015; Xu et al. 2013a, b, c, d, 2016a, b). Besides the Mann–Kendall trend test, we used a variety of methods to verify the trends, such as the Pettitt’s test (Pettitt 1979) to detect change-point, the Sen’s slope method (Sen 1968) to analyze trend. Seasonal Mann–Kendall test, correlated Seasonal Mann–Kendall test and seasonal Sen’s slope were applied to the monthly series to validate the annual trend test (Pohlert 2016). The steps of Mann–Kendall trend test are as follows:

For the time series \({X_t}=\left( {{x_1},{x_2}, \ldots ,{x_n}} \right)\), the statistics S of Mann–Kendall trend test was calculated with the following Eq. (13):
$$S=\mathop \sum \limits_{i=1}^{n - 1} \mathop \sum \limits_{k=i+1}^n sgn\left( {{x_k} - {x_i}} \right).$$
(13)
where, n is the length of the sample data, x i is the ith data in the time series and sgn is a sign function:
$$sgn\left( \theta \right)=\left\{ {\begin{array}{*{20}{c}} {1,\theta>0} \\ {0,\theta =0} \\ { - 1,\theta <0} \end{array}} \right..$$
(14)
If n is more equal than 8, the statistics S is approximately normally distributed. The average of the S equals to 0 and the variance was:
$$Var\left( s \right)=\frac{{n\left( {n - 1} \right)\left( {2n+5} \right) - \mathop \sum \nolimits_{i=1}^n {t_i}\left( {i - 1} \right)\left( {2i+5} \right)}}{{18}},$$
(15)
where, \({t_i}\) is the number of the data in group i.
The standardization statistics \({Z_c}\) was calculated as follows:
$${Z_c}=\left\{ {\begin{array}{*{20}{c}} {\frac{{S - 1}}{{\sqrt {Var\left( s \right)} }},S>0} \\ {0,S=0} \\ {\frac{{S+1}}{{\sqrt {Var\left( s \right)} }},S<0} \end{array}} \right..$$
(16)

If \({Z_c}\) is positive, the test series have an increasing trend, otherwise the trend is down. The null hypothesis is that the time series doesn’t vary significantly. If the absolute value of \({Z_c}\) is greater than 1.96 (95% confidence level), the time series varies significantly.

2.3.3 BP artificial neural network

The downscaled climate data made up for the lack of observed data. In order to quantitatively assess the impact of climate variability on streamflow, the third sub model used BPANN to simulate the nonlinear relationship among the monthly temperature t m , monthly precipitation p m and monthly streamflow s m .

BPANN can effectively simulate the complex nonlinear relationship between climate and hydrology (Xu et al. 2013a, b, c, d, 2014, 2016a, b). BPANN consists of an input layer, one or more hidden layers and an output layer. In this paper, the sub model adopted a neural network of 5 × 5 hidden layers structure by try and error (Fig. 4). In the neural network, t m and p m were the inputs to train the network, s m was the training target, w ij was the connection weight between the ith neuron in the input layer and the jth neuron in the hidden layer, w jk was the connection weight between the jth neuron in the hidden layer 1 and the kth neuron in the hidden layer 2, and w k was the connection weight between the lth neuron in the hidden layer and the lth neuron in the output layer.

Fig. 4

The structure of the BPANN

In the same layer, there is no connection among the neuron nodes. The output of a neuron node just connects to the input of the node in the next layer. The simulating process consists of the forward transmission of information and the back propagation of the error. In one training, the input and target data are divided into a training group (70%), a calibration group (15%) and a validation group (15%) randomly. When entering the neural network, the information is transferred from the input layer to the hidden layer at first and then transferred into the output layer by activation function. In this paper, the activation function was the sigmoid function. When the output didn’t match the target, the error was transferred back. The error was used to modify the connection weights by gradient decent algorithm from the output layer to the hidden layer and input layer. The neural network used the Levenberg–Marquardt algorithm to accelerate the convergence. After 1000 trainings, the neural network with the minimum square error was chosen as the optimal result of one BPANN fitting.

Because of the randomized initial parameters, each fitting result was different with others. Therefore the sub model repeated the fitting multiple times. Then the mean and the standard error of the outputs were calculated as the final result. In this research, the mean of the outputs was relatively stable after ten fittings after test.

The sub model calculated the Slope, R 2 , MAE and RMSE between the target s a and the mean of the outputs to verify the BPANN fitting results.

2.3.4 The impact of climate variability on streamflow

Compared with the traditional multiple regression analysis, the BPANN can improve the fitting accuracy, but the ability to explain the relationship among variables is weak. Therefore, we developed an improved weight connection method based on previous researches (Fischer 2015; Olden and Jackson 2002; Olden et al. 2004) to calculate the contribution of temperature and precipitation on streamflow.

At first, the third sub model calculated the cumulative weight matrix from the input layer to output layer as follows:
$${w_{io}}=\left[ {\begin{array}{*{20}{c}} {{w_{ts}}}&{{w_{ps}}} \end{array}} \right]={w_{{h_2}o}} \times {w_{{h_1}{h_2}}} \times {w_{i{h_1}}},$$
(17)
where \({w_{{h_2}o}}\) is the weight matrix (1 × 5) from the hidden layer 2 to the output layer, \({w_{{h_1}{h_2}}}\) is the weight matrix (5 × 5) from the hidden layer 1 to the hidden layer 2, \({w_{i{h_1}}}\) is the weight matrix (5 × 2) from the input layer to the hidden layer 1, \({w_{io}}\) is the cumulative weight matrix from the input layer to the output layer, \({w_{ts}}\) is the cumulative weight from the input of temperature to the output of streamflow, \(~{w_{ps}}\) is the cumulative weight matrix from input of precipitation to the output of streamflow.
Then the contribution could be calculated as follow equations:
$${C_{ts}}=\frac{{\left| {{w_{ts}}} \right|}}{{\left| {{w_{ts}}} \right|+\left| {{w_{ps}}} \right|}} \times 100{\% ,}$$
(18)
$${C_{ps}}=\frac{{\left| {{w_{ps}}} \right|}}{{\left| {{w_{ts}}} \right|+\left| {{w_{ps}}} \right|}} \times 100{\% ,}$$
(19)
where C ts is the contribution of temperature variability on stream variability, C ps is the contribution of precipitation variability on stream variability.

Because the randomized initial parameters of BPANN would lead to different fitting results, the third sub model calculated the mean and standard error of the contributions after ten fittings as the final result.

3 Results and discussion

3.1 Downscaling accuracy test

The fitting results of the temperature and precipitation with altitude were tested at first (Table 1). Based on analysis of all months, the linear relationship between temperature and altitude was strong (R 2 = 0.971) and significant (F = 2133.257, a = 0.005). The temperature declined steadily with the rising altitude at a rate of −0.006 °C m−1 according to the slope of the linear fitting function. The lapse rate was consistent with previous research results. For example, based on the average monthly temperature data of 460 meteorological stations from the northern Italy to the southern slope of the Alps, Rolland (2003) found that the temperature lapse rate was −0.0054 to −0.0058 °C m−1. The quadratic polynomial relationship between the precipitation and altitude was strong (R 2 = 0.648) and significant (F = 94.101, a = 0.005). The precipitation increased with the rising altitude at first, then gradually became stable at the altitude of about 3000 m and declined above 3000 m. The quadratic polynomial relationship was consistent with numerous studies (Chen et al. 2013). For instance, the relationship between precipitation and altitude was also quadratic polynomial in the Kaidu basin on the southern slope of the eastern Tianshan Mountains and the extreme point of precipitation was at 2450 m (Fu et al. 2013). We fitted the monthly climate data with the same method to show the spatiotemporal pattern of monthly temperature and precipitation. According to R 2 and F test, all the fittings were strong and significant (a = 0.005). The temperature lapse rate was large from March to October and relatively small from November to February next year. The precipitation increased first and then decreased with the increasing elevation. The fitting functions of temperature and precipitation with altitude can be applied in simulating the temperature lapse rates and precipitation gradients.

Table 1

The fitting functions of the temperature and precipitation with altitude

Time scales

N

Temperature

Precipitation

Fitting functions

R 2

F

Fitting functions

R 2

F

All months

1584

t = −0.006 × h + 18.200

0.971

2144.357*

p = −4.205E−6 × h 2 + 0.025 × h − 15.405

0.648

94.101*

January

132

t = −0.003 × h − 3.723

0.668

104.752*

p = −1.413E−7 × h 2 + 0.001 × h − 0.296

0.526

24.672*

February

132

t = −0.005 × h + 4.577

0.882

455.385*

p = −2.545E−6 × h 2 + 0.013 × h − 9.891

0.549

27.828*

March

132

t = −0.006 × h + 14.780

0.931

845.690*

p = −4.888E−6 × h 2 + 0.028 × h − 22.920

0.606

37.434*

April

132

t = −0.007 × h + 24.170

0.947

1129.800*

p = −6.488E−6 × h 2 + 0.036 × h − 31.140

0.623

40.914*

May

132

t = −0.008 × h + 29.174

0.959

1488.546*

p = −9.985E−6 × h 2 + 0.057 × h − 44.569

0.670

52.539*

June

132

t = −0.007 × h + 32.180

0.961

1569.813*

p = −4.147E−6 × h 2 + 0.030 × h −18.004

0.625

41.346*

July

132

t = −0.007 × h + 33.075

0.963

1659.869*

p = −1.462E−5 × h 2 + 0.077 × h − 53.764

0.600

36.281*

August

132

t = −0.007 × h + 31.849

0.950

1203.333*

p = −1.705E−5 × h 2 + 0.087 × h − 63.439

0.588

34.085*

September

132

t = −0.006 × h + 26.837

0.953

1286.261*

p = −1.775E−5 × h 2 + 0.088 × h − 70.730

0.595

35.349*

October

132

t = −0.006 × h + 18.648

0.931

845.690*

p = −1.355E−6 × h 2 + 0.011 × h − 9.597

0.621

40.487*

November

132

t = −0.004 × h + 8.097

0.884

464.846*

p = −5.612E−7 × h 2 + 0.005 × h − 4.312

0.564

30.088*

December

132

=−0.0003 × h − 1.263**

0.761

178.877*

p = −8.567E−8 × h 2 + 0.001 × h − 0.415

0.526

24.672*

F test indicate the significance of the fitting functions. *indicates the significance of a = 0.005

The downscaling method based on earth data products can simulate the mountainous climate well (Table 2). The stations for verification near the catchments were sufficient. The altitudes of these meteorological stations (Aksu, Keping, Aheqi and Chatyr Kul) varied from 1000 to 3500 m. The stations located at different terrains such as alluvial fan, hills and mountains. The verification was also guaranteed as the large number of observed monthly temperature and precipitation data. The slope was close to 1 and it meant the variety of the downscaled data was consistent with the observed data. The NSE was close to 1, which meant the simulation effect of the downscaling method based on earth data products was well. MAE and RMSE were both relatively small compared with the observed data, which meant the downscaled data were close to the observed data. The Figs. 5 and 6 showed that the effect of the downscaled data for temperature and precipitation was good. The observed and downscaled data in scatter plots concentrated near the trend line with a slope of 1. We also calculated the R 2 to validate the effect for downscaling. The all values of R 2 of the observed and downscaled data for monthly temperature and precipitation were close to 1. In summary, the downscaling method can be applied on the research of basin climate variability.

Table 2

Accuracy test of downscaling based on earth data products

Factors

Stations

N

Slope

NSE

MAE

RMSE

Temperature (°C)

Aksu

396

1.01

0.99

1.58

1.90

Keping

396

0.92

0.99

1.28

1.75

Aheqi

396

0.85

0.98

2.72

3.06

Chatyr Kul

396

0.96

0.99

0.95

1.19

Precipitation (mm)

Aksu

396

0.79

0.72

2.74

5.40

Keping

396

0.84

0.67

4.20

8.79

Aheqi

396

1.33

0.80

12.88

22.10

Chatyr Kul

396

0.93

0.95

2.93

5.00

The unit of MAE and RMSE is same as the original data

Fig. 5

The scatter plot and trend line of the downscaled and observed monthly temperature in verification stations. ad are scatter plots of the downscaled and observed monthly temperature in Aksu, Keping, Aheqi and Chatyr Kul, respectively

Fig. 6

The scatter plot and trend line of the downscaled and observed monthly precipitation in verification stations. ad are scatter plots of the downscaled and observed monthly precipitation in Aksu, Keping, Aheqi and Chatyr Kul, respectively

3.2 The temporal and spatial climate variability

Based on the descriptive statistics and Mann–Kendall trend test, the climate characteristics were different in the two catchments (Table 3). Although the average and the highest altitude of the Kumarik catchment were both higher than the Toshkan catchment, the average annual temperature in the Kumarik catchment from 1980 to 2012 was −0.81 °C, still higher than the −2.37 °C in the Toshkan catchment. The reason might be that the median of the altitude was respectively 3737 and 3768 m in the two catchments, and the lowest altitude of the Kumarik catchment was lower than the Toshkan catchment. The average annual precipitation in the Kumarik catchment during 32 years was 503.70 mm, which was higher than the 319.02 mm in the Toshkan catchment. The precipitation downscaling results were close to the model simulation results by Duethmann et al. (2015). He found the average annual precipitation in the Kumarik and Toshkan catchment from 1957 to 2004 was respectively 474 (450–526) mm and 386 (372–399) mm.

Table 3

Descriptive statistics and trend test of the annual temperature and precipitation from 1980 to 2012

Factors

Catchments

Descriptive statistics

Mann–Kendall trend test

Pettitt’s test

Sen’s slope

N

Mean

SD

CV

\(slope\)

Z

p

Change-point

Temperature (°C)

Kumarik

33

−0.81

0.46

−0.57

0.01

1.1

0.18

No

0.01

Toshkan

33

−2.37

0.46

−0.19

0.01

0.6

0.35

No

0.01

Precipitation (mm)

Kumarik

33

503.70

136.01

0.27

4.85

1.8*

0.34

No

4.17

Toshkan

33

319.02

95.43

0.30

4.10

2.4**

0.11

No

3.60

The unit of MEAN and SD is same as the original data. *Indicates the significance of a = 0.05. **Indicates the significance of a = 0.01

The annual differences of temperature in the Kumarik catchment were greater than the Toshkan catchment. It might be because the percent of the area above 4000 m was respectively 35 and 19% in the Kumarik and Toshkan catchment, and the higher the altitude, the higher temperature variability among years (Ohmura 2012). However, the annual differences of precipitation were similar in both catchments. It was possible related to the decreasing precipitation above 3500 m.

The Mann–Kendall trend test results showed that the average annual temperature in the Kumarik and Toshkan catchment from 1980 to 2012 both appeared an increasing trend (Z = 1.07; Z = 0.6) at the rate of 0.1 °C per decade. The annual precipitation in the Kumarik and Toshkan catchment showed a significant increasing trend (Z = 1.81, a = 0.05; Z = 2.40, a = 0.01) at the rate of 48.5 mm per decade and 41.0 mm per decade.

Sen’s slope test indicated the increasing trend of annual temperature and precipitation in the two catchments. The seasonal tests (Table 4) also showed the increasing trends of the monthly temperature and precipitation (Z > 0; slope > 0). The climate became warming and wetting in the headwaters of the Aksu River in recent 30 years, which was consistent with Xu’s (2006) research result.

Table 4

The seasonal test of the monthly temperature and precipitation from 1980 to 2012

Factors

Catchments

N

Seasonal Mann–Kendall test

Seasonal Mann–Kendall test with correlations

Seasonal Sen’s slope

Z

Z

Temperature (°C)

Kumarik

396

1.3*

1.1

0.01

Toshkan

396

0.8

0.7

0.01

Precipitation (mm)

Kumarik

396

0.8

0.8

0.02

Toshkan

396

1.5*

1.3*

0.04

*Indicates the significance of a = 0.1

We used the Pettitt’s test to detect change-point at a significant level of 0.05 (a = 0.05), and did not find obvious change-points of the time series data of temperature and precipitation. For the purpose of studying the inter-decadal relationship between the climate factors and streamflow (Bai et al. 2015), we divided the data series into four periods (i.e. 1980–1989, 1990–1999, and 2000–2012), and analyzed the nonlinear variation of climatic-hydrological processes in different periods.

The average value and the slope (rate of change) of downscaled (90 m × 90 m) grid annual temperature in different periods showed the spatial and temporal temperature variability (Figs. 7, 8). The temperature spatial distribution was similar in different periods. The temperature was generally above 0 °C in valleys and alluvial fan and below 0 °C in mountains. However, the temperature trend was different in each period. From 1980 to 1989, the increasing trend of temperature in the two catchments was relatively slow. The rising trend was close to 0 °C per decade in most regions except the eastern mountains of the Kumarik catchment. From 1990 to 2000, the temperature was increasing in both catchments. The rising trend was higher in alluvial fan than the mountains. The northeast mountains of the Toshkan catchment had the slowest rising trend. The rising trend in the Kumarik catchment was generally higher than the Toshkan catchment. From 2000 to 2012, the temperature was declining in both of the catchments. The decreasing trend was slow in the eastern mountains of the Kumarik catchment than other regions. From 1980 to 2012, the temperature in both catchments was increasing slowly. The rising trend in eastern Toshkan catchment was the slowest. The rising trend in the Kumarik catchment was generally higher than the Toshkan catchment.

Fig. 7

The average of downscaled annual temperature in different periods. a Is the average annual temperature from 1980 to 1989; b is the average annual temperature from 1990 to 1999; c is the average annual temperature from 2000 to 2012; d is the average annual temperature from 1980 to 2012

Fig. 8

The slope (rate of change) of downscaled annual temperature in different periods. a Is the slope of annual temperature from 1980 to 1989; b is the slope of annual temperature from 1990 to 1999; c is the slope of annual temperature from 2000 to 2012; d is the slope of annual temperature from 1980 to 2012

The temperature variability in the inland river basins of Northwest China was related to the Siberian high intensity and the carbon dioxide concentration (Chen et al. 2015). Because of the weakening of the Siberian high in recent 30 years, the frequency of the southward cold air was reducing and the strength was decreasing (Xu et al. 2006). So the temperature increasing trend decreased from the northern to the southern China (Yatagai and Yasunari 1994), and the increasing trend in the northern Kumarik catchment was generally higher than the southern Toshkan catchment (Xu et al. 2006). The increasing trend was fast in the alluvial fan because of the lower altitude and latitude. The increasing trend was slow in mountains, and it might be related to the buffering effect from the widespread glacier and snow, diverse vegetation and the stability of mountain ecosystems against the global climate variability (Li et al. 2013). The temperature trend in the headwaters of the Aksu River was consistent with the regular in Northwest China (Chen et al. 2015). The temperature in Northwest China was slowly increasing in the 1980s, then experienced an abrupt changing in 1987 and quickly increasing in the following 1990s. The temperature variability was related to the rapid increasing of population, industrialization, urbanization and greenhouse gas emissions (Li et al. 2013).

The average value and the slope (rate of change) of downscaled (90 m × 90 m) grid annual precipitation in different periods showed the spatial and temporal precipitation variability (Figs. 9, 10). The precipitation spatial distribution was similar in different periods. The precipitation decreased from the northeast to southwest. As a result, the precipitation was the most abundant in the mountains of the Kumarik catchment and the least in the west and southern mountains of the Toshkan catchment. The precipitation trend was different in each period. From 1980 to 1989, the precipitation increasing trend decreased from west to east and became decreasing in the Kumarik catchment. From 1990 to 1999, the precipitation in the two catchments was increasing. Only in the alluvial fan of the Kumarik catchment and low altitude valleys of the Toshkan catchment the rising trend was slightly slow. From 2000 to 2012, the precipitation in the two catchments still kept rising. From 1980 to 2012, the precipitation was increasing in both of the catchments and the rising trend in the Kumarik catchment was greater than the Toshkan catchment.

Fig. 9

The average of downscaled annual precipitation in different periods. a Is the average annual precipitation from 1980 to 1989; b is the average annual precipitation from 1990 to 1999; c is the average annual precipitation from 2000 to 2012; d is the average annual precipitation from 1980 to 2012

Fig. 10

The slope (rate of change) of downscaled annual precipitation in different periods. a Is the slope of annual precipitation from 1980 to 1989; b is the slope of annual precipitation from 1990 to 1999; c is the slope of annual precipitation from 2000 to 2012; d is the slope of annual precipitation from 1980 to 2012

The precipitation in the Kumarik catchment was greater than the Toshkan catchment. It was because the glacier and snow were widely distributed in the Kumarik catchment, and the agricultural activity and evaporation were much in the city of Aksu near the outlet of the Kumarik catchment (Xu et al. 2006). So the Kumarik catchment was wetter than the Toshkan catchment. That the increasing trend in mountains was higher than the alluvial fan and valleys was consistent with the Fu et al.’s (2013) research. It was because the mountainous terrain was advantageous to formatting precipitation, together with the increasing water vapor caused by warming climate and melting glacier and snow (Yang et al. 2009). The decreasing precipitation in 1980s verified the previous research (Chen et al. 2013; Li et al. 2013). The rapid increasing of precipitation in 1990s and recent years was related to the climate warming and the acceleration of water vapor cycle (Chen et al. 2013). The trend of temperature and precipitation in different periods might be related to the climate fluctuations and ENSO events (Infanti and Kirtman 2016). Tree-ring studies showed that Northwest China experienced cyclical change between warm and cold together with wet and dry in the past 200–300 years (Chen et al. 2012; Yang et al. 2012).

3.3 The impact of temperature and precipitation on streamflow

The annual streamflow increased in both of the Kumarik catchment and Toshkan catchment from 1980 to 2012 (Table 5). The increasing trend was related to the rising temperature and precipitation in the two catchments and consistent with Xu et al.’s (2006) research. The average annual streamflow of the Kumarik River and Toshkan River in 33 years was respectively 51.42 × 108 and 29.87 × 108 m3. Because of the more precipitation and glaciers, the streamflow in the Kumarik catchment was greater than the Toshkan catchment. From the SD and CV, the annual difference of streamflow in the Kumarik catchment was greater than the Toshkan River. It might be related to the annual melting difference of glacier and snow caused by temperature variability. The Mann–Kandall test shows that the streamflow in the Kumarik and Toshkan catchment were both increasing significantly (Z = 1.47, a = 0.1; Z = 3.27, a = 0.01) at a rate of 1.5 × 108 and 3.3 × 108 m3 per decade respectively. Sen’s slope test indicated the increasing trend of annual streamflow in the two catchments. The seasonal tests (Table 6) also showed the increasing trends of the monthly streamflow (Z > 0; slope > 0) and the trends were significant (a = 0.01). The Pettitt’s test detected the change-points in year 1994.

Table 5

Descriptive statistics and trend test of the annual streamflow from 1980 to 2012

Rivers

N

Descriptive statistics

Mann–Kendall trend test

Pettitt’s test

Sen’s slope

Mean (m3 s−1)

SD

CV

\(slope\)

Z

p

Change-point

Kumarik

33

51.42

7.14

0.14

0.15

1.47*

0.004

1994

0.16

Toshkan

33

29.87

5.76

0.19

0.33

3.27**

0.002

1994

0.35

*Indicates the significance of a = 0.1. **Indicates the significance of a = 0.01

Table 6

The seasonal test of the monthly streamflow from 1980 to 2012

Rivers

N

Seasonal Mann–Kendall test

Seasonal Mann–Kendall test with correlations

Seasonal Sen’s slope

Z

Z

Kumarik

396

6.8*

3.9*

0.28

Toshkan

396

10.1*

4.6*

0.54

*Indicates the significance of a = 0.01

Table 7 shows the change trends of climate on the streamflow by comparing the slope of temperature, precipitation and streamflow in catchment based on downscaled data. In the 1980s, the streamflow and the precipitation both decreased while the temperature was relatively stable in the Kuamrik catchment. In the same period, the streamflow and temperature kept relatively stable while the precipitation was slightly decreasing in the Toshkan catchment. In the 1990s, the streamflow of the Kuamrik and Toshkan river both increased with the rising temperature and precipitation. It was consistent with Chen et al.’s (2013) research that the Tarim River basin was warmest and wettest and the streamflow was increasing fastest in the 1990s. Because of the faster rising rates of temperature and precipitation, the streamflow increasing trend in the Kumarik catchment was also faster than the Toshkan catchment. At the beginning of twenty-first century, although the precipitation was increasing, the streamflow in both catchments decreased with the declining temperature. The decreasing trend of the streamflow in the Kumarik catchment was faster than the Toshkan catchment. It might be because the Kumarik catchment was more sensitive to the temperature variability.

Table 7

The slope of annual temperature, precipitation and streamflow in different periods

Factors

1980–1989

1990–1999

2000–2012

Kumarik

Toshkan

Kumarik

Toshkan

Kumarik

Toshkan

Temperature

−0.01

−0.01

0.04

0.03

−0.03

−0.04

Precipitation

−17.92

−5.02

12.51

9.30

11.23

10.03

Streamflow

−0.49

0.04

2.07

1.00

−0.86

−0.42

Based on the idea and framework of the hybrid model mentioned previously in the methodology of this study, we computed the simulation values of streamflow by the BPANN based on downscaled temperature and precipitation at the monthly scale (Fig. 11).

Fig. 11

The monthly streamflow observed and simulated with BPANN. a Is the monthly streamflow observed and simulated in the Kumarik river; b is the monthly streamflow observed and simulated in the Toshkan river

In order to compare and validate the simulated results from the hybrid model, we also simulated the monthly streamflow with the monthly temperature and precipitation by the classical approach, i.e. the multiple regression. Tables 8 and 9 reveal the compared results. It evident that the simulated precise of streamflow by the BPANN based on downscaled climate data is higher than the simulated results by the multiple regression.

Table 8

The multiple regression fitting functions and accuracy test of the monthly temperature, precipitation and streamflow

Rivers

Fitting functions

Accuracy test

P00

P10

P01

P20

P11

P02

N

Slope

NSE

MAE

RMSE

Kumarik

76.36

21.53

−0.59

1.70

0.08

0.002421

396

0.88

0.88

48.20

65.34

Toshkan

70.02

9.36

0.61

0.41

0.02

0.000745

396

0.76

0.76

26.59

43.79

The fitting function is s = p00 + p10 × t + p01 × p + p20 × t^2 + p11 × t × p + p02 × p^2. Variable s is the monthly streamflow, t is the monthly temperature and p is the monthly precipitation

Table 9

The accuracy test of BPANN and the contribution assessing

Rivers

Accuracy test

Contribution (%)

N

Slope

NSE

MAE

RMSE

Temperature

Precipitation

Kumarik

396

0.94

0.95

25.24

42.80

64.01 ± 7.34

35.99 ± 7.34

Toshkan

396

0.81

0.83

22.39

37.15

47.74 ± 8.10

52.26 ± 8.10

The input data of the BPANN are monthly temperature and precipitation. The target data is the monthly streamflow. The unit of MAE and RMSE is same as the original data (m3 s−1)

Figure 11 showed that the BPANN well simulated the monthly streamflow with the monthly temperature and precipitation. The slope was close to 1 and indicated that the simulated data was consistent with the observed data. The NSE was close to 1 which illustrated the BPANN with a hidden layer of 5 × 5 could simulate the streamflow well after ten fittings. The MAE and the RMSE are both small which represented the high precision of the simulation results (Table 9). We calculated the simulated annual streamflow by add up monthly values. The simulated annual streamflow was consistent with the observed series (Fig. 12).

Fig. 12

The annual streamflow observed and simulated with BPANN. a Is the annual streamflow observed and simulated in the Kumarik river; b is the streamflow observed and simulated in the Toshkan river

In summary, the BPANN could simulate the nonlinear relationship among the monthly and annual temperature, precipitation and streamflow. The fitting results could guarantee the high accuracy of quantitatively assessing the impact of the climate variability on streamflow.

Based on the monthly BPANN and weight connection method, the impact of the temperature and precipitation on streamflow could be assessed quantitatively (Table 9). The contribution of the temperature and precipitation on the streamflow was respectively 64.01 ± 7.34 and 35.99 ± 7.34% in the Kumarik catchment, and the contribution was 47.74 ± 8.10 and 52.26 ± 8.10% in the Toshkan catchment. The Kumarik catchment was mainly impacted by temperature while the Toshkan catchment was mainly impacted by precipitation. The results were consistent with the analysis based on the Table 7 and Duethmann et al.’s (2015) research that the Kumarik River was mainly supplied with glacier and snow melting water and the Toshkan River was mainly supplied with the precipitation.

Climate variability leads to the change of the regional water resources (Liu et al. 2016). Especially in arid and semi-arid areas, the slight change of the temperature and precipitation would result in a great effect on the streamflow (Gan 2000). Xu et al.’s (2006) research shows that in the Aksu River, the glacier and snow melting water and the direct precipitation were the main water supply. So the streamflow increased with the rising temperature and precipitation from 1980 to 2012. Further more, the impact of climate on streamflow was affected by the underlying of the catchments (Zhang and Zuo 2015). The Kumarik catchment was covered with more glacier and snow than the Toshkan catchment, so the contribution of temperature would be greater in the Kuamrik catchment.

Based on the fitting functions of the temperature and precipitation with altitude, we downscaled the earth data products and simulated the mountainous climate changes. The climate became warming and wetting in the headwaters of the Aksu River in recent 30 years. The annual streamflow increased in both of the Kumarik catchment and Toshkan catchment. The BPANN can simulate the annual streamflow with the annual temperature and precipitation well. Based on the BPANN and weight connection method, the impact of the temperature and precipitation on streamflow could be assessed quantitatively. The hybrid model revealed regional climate change characteristics and achieved the research objectives.

4 Conclusions

To quantitatively assess the impact of climate variability on streamflow for an ungauged mountainous basin, we conducted a hybrid model by integrating a downscaling method based on earth data products, Mann–Kendall trend test, back propagation artificial neural networks (BPANN) and weights connection method, and validated the model by the mountainous basin of the Aksu River in Northwest China. The conclusion is that the hybrid model presented a good performance, and the quantitative assessment results for the mountainous basin of the Aksu River are as follows:

  1. 1.

    The precipitation respectively increased by 48.5 and 41.0 mm per decade in the Kumarik catchment and Toshkan catchment during the period from 1980 to 2012, and the average annual temperature both increased by 0.1 °C per decade in the two catchments.

     
  2. 2.

    With the warming and wetting climate, the streamflow respectively increased 1.5 × 108 and 3.3 × 108 m3 per decade in the Kumarik River and the Toshkan River.

     
  3. 3.

    The contribution of the temperature and precipitation on the streamflow was respectively 64.01 ± 7.34 and 35.99 ± 7.34% in the Kumarik catchment, and that was 47.74 ± 8.10 and 52.26 ± 8.10% in the Toshkan catchment.

     

Our study introduced a feasible hybrid model for the assessment of the impact of climate variability on streamflow, which can be used in the ungauged mountainous basin of Northwest China.

Notes

Acknowledgements

This work was supported by the Open Foundation of the State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences (No. G2014-02-07); and the National Natural Science Foundation of China (41630859).

References

  1. Allen MR, Frame DJ, Huntingford C, Jones CD, Lowe JA (2009) Warming caused by cumulative carbon emissions towards the trillionth tonne. Nature 458:1163–1166CrossRefGoogle Scholar
  2. Bai L, Chen Z, Xu J, Li W (2015) Multi-scale response of runoff to climate fluctuation in the headwater region of Kaidu River in Xinjiang of China. Theor Appl Climatol 2016:1–10. doi: 10.1007/s00704-015-1539-2 Google Scholar
  3. Brown ME, Racoviteanu AE, Tarboton DG et al (2014) An integrated modeling system for estimating glacier and snow melt driven streamflow from remote sensing and earth system data products in the Himalayas. J Hydrol 519:1859–1869. doi: 10.1016/j.jhydrol.2014.09.050 CrossRefGoogle Scholar
  4. Chen Y, Cui W, Li W, Zhang Y (2003) Utilization of water resources and ecological protection in the Tarim River. Acta Geographica Sinica 58:215–222. doi: 10.11821/xb200302008 Google Scholar
  5. Chen Y, Takeuchi K, Xu C, Chen Y, Xu Z (2006) Regional climate change and its effects on river runoff in the Tarim Basin, China. Hydrol Process 20:2207–2216. doi: 10.1002/hyp.6200 CrossRefGoogle Scholar
  6. Chen F, Yuan Y, Wen W et al (2012) Tree-ring-based reconstruction of precipitation in the Changling Mountains, China, since A.D.1691. Int J Biometeorol 56:765–774. doi: 10.1007/s00484-011-0431-8 CrossRefGoogle Scholar
  7. Chen Y, Xu C, Chen Y, Liu Y, Li W (2013) Progress, challenges and prospects of eco-hydrological studies in the Tarim River Basin of Xinjiang, China. Environ Manage 51:138–153. doi: 10.1007/s00267-012-9823-8 CrossRefGoogle Scholar
  8. Chen Y, Li Z, Fan Y, Wang H, Deng H (2015) Progress and prospects of climate change impacts on hydrology in the arid region of northwest China. Environ Res 139:11–19. doi: 10.1016/j.envres.2014.12.029 CrossRefGoogle Scholar
  9. Duethmann D, Bolch T, Farinotti D et al (2015) Attribution of streamflow trends in snow and glacier melt-dominated catchments of the Tarim River, Central Asia. Water Resour Res 51:4727–4750. doi: 10.1002/2014WR016716 CrossRefGoogle Scholar
  10. Fischer A (2015) How to determine the unique contributions of input-variables to the nonlinear regression function of a multilayer perceptron. Ecol Modell 2015:60–63CrossRefGoogle Scholar
  11. Fu AH, Chen YN, Li WH, Li BF, Yang YH, Zhang SH (2013) Spatial and temporal patterns of climate variations in the Kaidu River Basin of Xinjiang, Northwest China. Quat Int 311:117–122. doi: 10.1016/j.quaint.2013.08.041 CrossRefGoogle Scholar
  12. Gan TY (2000) Reducing vulnerability of water resources of Canadian prairies to potential droughts and possible climatic warming. Water Resour Manag 14:111–135CrossRefGoogle Scholar
  13. Georgakakos KP, Graham NE, Modrick TM, Murphy MJ, Shamir E, Spencer CR, Sperfslage JA (2014) Evaluation of real-time hydrometeorological ensemble prediction on hydrologic scales in Northern California. J Hydrol 519:2978–3000. doi: 10.1016/j.jhydrol.2014.05.032 CrossRefGoogle Scholar
  14. Getirana ACV, Espinoza JCV, Ronchail J, Rotunno Filho OC (2011) Assessment of different precipitation datasets and their impacts on the water balance of the Negro River basin. J Hydrol 404:304–322. doi: 10.1016/j.jhydrol.2011.04.037 CrossRefGoogle Scholar
  15. Guo Y, Shen Y (2016) Agricultural water supply/demand changes under projected future climate change in the arid region of northwestern China. J Hydrol 540:257–273. doi: 10.1016/j.jhydrol.2016.06.033 CrossRefGoogle Scholar
  16. Infanti JM, Kirtman BP (2016) North American rainfall and temperature prediction response to the diversity of ENSO. Clim Dyn 46:3007–3023. doi: 10.1007/s00382-015-2749-0 CrossRefGoogle Scholar
  17. Kendall MG (1948) Rank correlation methods. Oxford Univ Pr, EnglandGoogle Scholar
  18. Li B, Chen Y, Shi X, Chen Z, Li W (2013) Temperature and precipitation changes in different environments in the arid region of northwest China. Theor Appl Climatol 112:589–596. doi: 10.1007/s00704-012-0753-4 CrossRefGoogle Scholar
  19. Liston GE, Elder K (2006) A meteorological distribution system for high-resolution terrestrial modeling (MicroMet). J Hydrometeorol 7:217–234. doi: 10.1175/JHM486.1 CrossRefGoogle Scholar
  20. Liu Z, Xu Z, Charles SP, Fu G, Liu L (2011) Evaluation of two statistical downscaling models for daily precipitation over an arid basin in China. Int J Climatol 31:2006–2020. doi: 10.1002/joc.2211 CrossRefGoogle Scholar
  21. Liu Y, Yang W, Qin C, Zhu A (2016) A review and discussion on modeling and assessing agricultural best management practices under global climate change. J Sustain Dev 9:245CrossRefGoogle Scholar
  22. Ma Z, Kang S, Zhang L, Tong L, Su X (2008) Analysis of impacts of climate variability and human activity on streamflow for a river basin in arid region of northwest China. J Hydrol 352:239–249. doi: 10.1016/j.jhydrol.2007.12.022 CrossRefGoogle Scholar
  23. Mahmood R, Babel MS (2013) Evaluation of SDSM developed by annual and monthly sub-models for downscaling temperature and precipitation in the Jhelum basin, Pakistan and India. Theor Appl Climatol 113:27–44. doi: 10.1007/s00704-012-0765-0 CrossRefGoogle Scholar
  24. Mann HB (1945) Nonparametric tests against trend. Econometrica J Econom Soc 1945:245–259CrossRefGoogle Scholar
  25. Neupane RP, White JD, Alexander SE (2015) Projected hydrologic changes in monsoon-dominated Himalaya Mountain basins with changing climate and deforestation. J Hydrol 525:216–230. doi: 10.1016/j.jhydrol.2015.03.048 CrossRefGoogle Scholar
  26. Ohmura A (2012) Enhanced temperature variability in high-altitude climate change. Theor Appl Climatol 110:499–508. doi: 10.1007/s00704-012-0687-x CrossRefGoogle Scholar
  27. Olden JD, Jackson DA (2002) Illuminating the “black box”: a randomization approach for understanding variable contributions in artificial neural networks. Ecol Modell 154:135–150. doi: 10.1016/S0304-3800(02)00064-9 CrossRefGoogle Scholar
  28. Olden JD, Joy MK, Death RG (2004) An accurate comparison of methods for quantifying variable importance in artificial neural networks using simulated data. Ecol Modell 178:389–397CrossRefGoogle Scholar
  29. Pachauri RK, Allen MR, Barros VR et al (2014) Climate change 2014: synthesis report. Contribution of Working Groups I, II and III to the fifth assessment report of the Intergovernmental Panel on Climate Change. IPCC, SwitzerlandGoogle Scholar
  30. Pettitt AN (1979) A non-parametric approach to the change-point problem. J R Stat Soc 28:126–135Google Scholar
  31. Pohlert T (2016) Non-Parametric Trend Tests and Change-Point Detection. CC BY-ND 4.0. http://creativecommons.org/licenses/by-nd/4.0/. Accessed 28 May 2017
  32. Raup B, Racoviteanu A, Khalsa SJS, Helm C, Armstrong R, Arnaud Y (2007) The GLIMS geospatial glacier database: a new tool for studying glacier change. Glob Planet Change 56:101–110. doi: 10.1016/j.gloplacha.2006.07.018 CrossRefGoogle Scholar
  33. Reuter HI, Nelson A, Jarvis A (2007) An evaluation of void-filling interpolation methods for SRTM data. Int J Geogr Inf Sci 21:983–1008CrossRefGoogle Scholar
  34. Rolland C (2003) Spatial and seasonal variations of air temperature lapse rates in alpine regions. J Clim 16:1032–1046CrossRefGoogle Scholar
  35. Sen PK (1968) Estimates of the regression coefficient based on Kendall’s Tau. J Am Stat Assoc 63:1379–1389CrossRefGoogle Scholar
  36. Xu C, Chen Y, Li W, Chen Y (2006) Climate change and hydrologic process response in the Tarim River Basin over the past 50 years. Chin Sci Bull 51:25–36. doi: 10.1007/s11434-006-8204-1 CrossRefGoogle Scholar
  37. Xu C, Chen Y, Hamid Y, Tashpolat T, Chen Y, Ge H, Li W (2009) Long-term change of seasonal snow cover and its effects on river runoff in the Tarim River basin, northwestern China. Hydrol Process 23:2045–2055. doi: 10.1002/hyp.7334 CrossRefGoogle Scholar
  38. Xu J, Chen Y, Li W et al (2013a) Combining BPANN and wavelet analysis to simulate hydro-climatic processes—a case study of the Kaidu River, North-west China. Front Earth Sci 7:227–237. doi: 10.1007/s11707-013-0354-2 CrossRefGoogle Scholar
  39. Xu J, Chen Y, Li W, Nie Q, Hong Y, Yang Y (2013b) The nonlinear hydro-climatic process in the Yarkand River, northwestern China. Stoch Environ Res Risk Assess 27:389–399. doi: 10.1007/s00477-012-0606-9 CrossRefGoogle Scholar
  40. Xu Z, Liu P, Liu W (2013c) Automated statistical downscaling in several river basins of the Eastern Monsoon region, China. IAHS AISH Publ 2013:81–85Google Scholar
  41. Xu C, Chen Y, Chen Y, Zhao R, Ding H (2013d) Responses of surface runoff to climate change and human activities in the arid region of Central Asia: a case study in the Tarim River Basin, China. Environ Manage 51:926–938. doi: 10.1007/s00267-013-0018-8 CrossRefGoogle Scholar
  42. Xu J, Chen Y, Li W, Nie Q, Song C, Wei C (2014) Integrating wavelet analysis and BPANN to simulate the annual runoff with regional climate change: a case study of Yarkand River, Northwest China. Water Resour Manag 28:2523–2537. doi: 10.1007/s11269-014-0625-z CrossRefGoogle Scholar
  43. Xu J, Chen Y, Bai L, Xu Y (2016a) A hybrid model to simulate the annual runoff of the Kaidu River in northwest China. Hydrol Earth Syst Sci 20:1447–1457. doi: 10.5194/hess-20-1447-2016 CrossRefGoogle Scholar
  44. Xu J, Chen Y, Li W, Liu Z, Tang J, Wei C (2016b) Understanding temporal and spatial complexity of precipitation distribution in Xinjiang, China. Theor Appl Climatol 123:321–333. doi: 10.1007/s00704-014-1364-z CrossRefGoogle Scholar
  45. Yang YH, Li WH, Wei WS, Hao XM, WAN M, LI H (2009) Discrepancy analysis of the climate changes among mountain, plain, oasis and desert in an inland river basin in the northern slopes of the Tianshan Mountains—a case study in the Sangong river basin. J Glaciol Geocryol 31:1094–1100Google Scholar
  46. Yang Y, Chen Y, Li W, Yu S, Wang M (2012) Climatic change of inland river basin in an arid area: a case study in northern Xinjiang, China. Theor Appl Climatol 107:143–154. doi: 10.1007/s00704-011-0467-z CrossRefGoogle Scholar
  47. Yang D, Gao B, Jiao Y, Lei H, Zhang Y, Yang H, Cong Z (2015) A distributed scheme developed for eco-hydrological modeling in the upper Heihe River. Sci China Earth Sci 58:36–45. doi: 10.1007/s11430-014-5029-7 CrossRefGoogle Scholar
  48. Yatagai A, Yasunari T (1994) Trends and decadal-scale fluctuations of surface air temperature and precipitation over China and Mongolia during the recent 40 year period (1951–1990). J Meteorol Soc Jpn 72:937–957CrossRefGoogle Scholar
  49. Zhang X, Zuo Q (2015) Analysis of water resource situation of the Tarim River basin and the system evolution under the changing environment. J Coastal Res 73:9–16. doi: 10.2112/SI73-003.1 CrossRefGoogle Scholar
  50. Zhao J, Xu Z, Singh VP (2016) Estimation of root zone storage capacity at the catchment scale using improved Mass Curve Technique. J Hydrol 540:959–972. doi: 10.1016/j.jhydrol.2016.07.013 CrossRefGoogle Scholar

Copyright information

© Springer-Verlag GmbH Germany 2017

Authors and Affiliations

  • Chong Wang
    • 1
    • 2
    • 3
  • Jianhua Xu
    • 1
    • 2
    • 3
  • Yaning Chen
    • 4
  • Ling Bai
    • 5
  • Zhongsheng Chen
    • 6
  1. 1.Key Laboratory of Geographic Information Science (Ministry of Education)East China Normal UniversityShanghaiChina
  2. 2.Research Center for East-West Cooperation in ChinaEast China Normal UniversityShanghaiChina
  3. 3.School of Geographic SciencesEast China Normal UniversityShanghaiChina
  4. 4.State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and GeographyChinese Academy of SciencesUrumqiChina
  5. 5.School of Economics and ManagementNanchang UniversityNanchangChina
  6. 6.College of Land and ResourcesChina West Normal UniversityNanchongChina

Personalised recommendations