Impact of Climate Variabilities and Human Activities on Surface Water Extents in Reservoirs of Yongding River Basin, China, from 1985 to 2016 Based on Landsat Observations and Time Series Analysis

Abstract

Yongding River is the largest river flowing through Beijing, the capital city of China. In recent years, Yongding River Basin (YDRB) has witnessed increasing human impacts on water resources, posing serious challenges in hydrological and ecological health. In this study, remote sensing techniques and statistical time series approaches for hydrological studies were combined to characterize the dynamics and driving factors of reservoir water extents in YDRB during 1985–2016. First, 107 Landsat 4, 5, 7 and 8 images were used to extract surface water extents in YDRB during 1985–2016 using a combination of water indices and Otsu threshold algorithm. Significant positive correlation was found between water extents and the annual inflow for the two biggest reservoirs, the downstream Guanting and upstream Cetian reservoirs, proving their representativeness of surface water availability in this basin. Then, statistical time series approaches including trend-free pre-whitening Mann-Kendall trend test, Pettit change-point test and double mass curve method, which are frequently used in hydrological studies, were adopted to quantify the trend of reservoir water extents dynamics and the relative contributions of climate variability and human activities. Results showed that the water extents in both reservoirs exhibited significant downward trend with change point occurring in 2001 and 2005 for Guanting and Cetian, respectively. About 74%~75% of the shrinkage during the post-change period can be attributed to human activities, among which GDP, population, electricity power production, raw coal production, steel and crude iron production, value of agriculture output, and urban area were the major human drivers. Hydrological connectivity between the upstream Cetian and downstream Guanting reservoirs declined during the post-change period. Since 2012, water extents in both reservoirs recovered as a result of various governmental water management policies including the South-to-North Water Diversion Project. The methodology presented in this study can be used for analyzing the dynamics and driving mechanism of surface water resources, especially for un-gauged or poorly-gauged watersheds.

1. Introduction

Water-related issues, including floods, droughts and water shortages have been identified as major global challenges in the 21st century [1]. In particular, the conflict between freshwater supply and demand has been widely recognized to be a source of strategic rivalry nationally, regionally and locally. While climate change might alter the hydrological cycle and the availability of water resources, human activities and socioeconomic development increasingly put pressure on freshwater supply, and may lead to water scarcity [2].

Quantifying the spatiotemporal dynamics of surface water resources and understanding their driving mechanism over the past decades will provide decision-makers with information for feasible restoration and management strategies and to further to evaluate their effects. Existing research in the hydrology community has generally utilized in-situ streamflow data obtained from hydrological gauge stations to analyze the temporal dynamics of surface water resources [3,4,5]. Compared to gauge station observations, remote sensing techniques provide direct and synoptic observations on spatiotemporal dynamics of surface water extents at regional and even global scales that can be used as indicators of water resources availability. In recent years, Landsat-series satellite datasets have been increasingly used to monitor surface water dynamics, given the advantage of their public accessibility, high resolution, and long-term availability (i.e., featuring a three-decade long data archive). Using these datasets, global open-water products have been developed, such as the Global 3 arc-second Water Body Map (G3WBM) at a five-year interval during 1990–2010 [6] at ~90 m resolution at the equator (http://hydro.iis.u-tokyo.ac.jp/~yamadai/G3WBM/), the 30 m-resolution Global Surface Water Explorer (GSWE) dataset from 1984 to 2015 [7] (https://global-surface-water.appspot.com/), and the 90m-resolution Global Inundation Extent from Multi-Satellite (GIEMS-D3) by combining Landsat-series and Synthetic Aperture Radar (SAR) datasets during the period of 1993–2007 [8,9]. Although these products provide good descriptions of surface water dynamics with regard to global coverage, they may not satisfy all the needs for regional-scale studies. Specifically, G3WBM and GIEMS-D3 have relatively short temporal coverages; GSWE do not provide annual time series of surface water extents, and only maximum water extent and water occurrence frequency were available. For time series analysis, annual or even seasonal water extents need to be extracted. Several remote-sensing-based indices are available for extracting water surfaces, including the Normalized Difference Water Index (NDWI) [10], the Modified Normalized Difference Water Index [11], the water index developed by Danaher and Collett (2006) [12], the Automated Water Extraction Index (AWEI) [13] and the water index developed by Fisher et al. (2016) [14]. All these indices allow water pixels to be identified by specifying a simple threshold, which can be adjusted for specific images. Compared to supervised classification methods, the water index method is easy to be implemented and facilitates fully automated water extraction.

In this study, we chose Yonding River Basin (YDRB) as our study region. Yongding River, the largest tributary of the Haihe River Basin (HRB), is commonly known as the mother river of the city of Beijing, China. The river basin is situated in a typical temperate semi-arid continental climate with long-term annual mean precipitation around 400 mm. The river flows through Shanxi Province, Inner Mongolia Autonomous Region, Hebei Province, and down to Beijing and Tianjin. Historically, wood harvest in the headwater region of the basin led to massive and frequent flood disasters prior to the 20th century [15]. For flood mitigation, dams and various water conservancy facilities were built in the YDRB, such as Guanting Reservoir built in 1954 and Cetian Reservoir built in 1960 [16]. In addition to flood control, these reservoirs have also played important roles in electricity power generation, and agricultural, domestic, and industrial supply. Over the past three decades, YDRB has become a region with intensive human activity. In the upper reach of YDRB, Shanxi Province produces about one-third of coal in China; Hebei Province is an important industrial base for steel and iron production [17]. Beijing, in the lower reach of YDRB, has been seeing massive developments at an unprecedented pace with rapid growth of the population and urban expansion. Increasing water abstraction has put more pressure on resources, causing serious water scarcity [1,3]. Water shortage has not only led to degradation of water quality and ecological security, hindering sustainable development, but it has also been reported that the lower reaches (the Beijing-Tianjin section) of Yongding River have been perennially dry [4]. In recent years, the state and local governments have implemented several inter-basin water transfer projects aiming to alleviate the water shortage in north China, including the South-to-North Water Diversion Project transporting water from the Yangtze River basin to north China and the Yellow River Diversion Project conveying water from the Yellow River to Shanxi Province. Intra-basin water transfer projects have also been implemented in YDRB, including water conveyance from upstream reservoirs to downstream reservoirs. Nevertheless, the degradation of aquatic ecosystems has not been fully mitigated. In 2017, the Development and Reform Commission of China announced the “The General Plan of Comprehensive Treatment and Ecological Restoration for the Yongding River Basin”, a cross-provincial program designed to restore riverine ecosystems in five to ten years.

The objectives of this study therefore are to: (1) investigate the annual dynamics and statistical trends of surface water extents in YDRB over the past 32 years (1985–2016) using Landsat-based datasets. Due to the narrow river channel width in YDRB from Landsat-series imagery, we hence mainly focus on water extents in reservoirs, as they represent the majority of the surface water storage in YDRB; (2) quantify the contribution of climate variability and human activities to variations in surface water extents in key reservoirs at the annual time scale; and (3) assess natural, socioeconomic factors and governmental policies that drive the temporal trends of reservoir water extents.

2. Study Area

The Yongding River (112°~117°45′ E, 39°~41°20′ N) is a major tributary of the Haihe River in North China (Figure 1). It is 747 km long, flowing through Shanxi Province, Inner Mongolia Autonomous Region, Hebei Province, Beijing and Tianjin, and drains into the Bohai Bay. Yanghe and Sanggan rivers are the two major tributaries of the Yongding River, originating from the Inner Mongolia Autonomous Region and Shanxi Province, respectively. Both rivers flow through the Guanting reservoir. The YDRB is located in a semi-humid and semi-arid climate transition zone with a drainage area of 47,016 km², annual average temperatures ranging between 5.1–7.1 °C and a mean annual precipitation of ~400 mm. The climate of YDRB features a strong seasonality with 70–80% of its rainfall occurring in the summer (i.e., June to August).

As aforementioned, rapid industrial development, population growth and urban expansion have led to intensive water abstraction during the past three decades. Water management policies have been implemented in recent years to alleviate the water shortage. In this study, we focus our analyses on the Guanting reservoir and its upstream drainage area. Currently, 14 large and medium reservoirs exist in the study domain, all of which were constructed between 1950 and 1980 [16]. These reservoirs have been mostly used for flood control and water supply, while some of them are designed for electricity generation [18]. Among these reservoirs, the Guanting reservoir in Beijing has the largest area, followed by the Cetian reservoir located along the Sangan river in Shanxi Province. Sedimentation in both reservoirs is marginal, according to regional water authority bureaus. Two major prefecture-level cities upstream to the Guanting reservoir are Datong in Shanxi Province and Zhangjiakou in Hebei Province. Major land cover types in the study area are forest (13%), grassland (26%), cropland (56%) and urban (5%).

3. Materials and Methods

3.1. Remote Sensing Imagery and Additional Datasets

In this study, a total of 107 images from Landsat 4 Multispectral Scanner (MSS) at 60-m resolution (1 scene), Landsat 4-5 Thematic Mapper (TM) (87 scenes), Landsat Enhanced Thematic Mapper (ETM+) (8 scenes) and Landsat 8 Operational Land Imager (OLI) images (11 scenes) at 30-m resolution acquired between 1985 and 2016 were used to monitor the dynamics of reservoir water extent in the YDRB (Figure A1). In order to minimize the impact of seasonal variability, only images acquired from late-August to October were selected because the water extents reach the maximum in a year as a result of summer precipitation. All images are cloudless over the reservoirs. Landsat 4 MSS level-1 terrain-corrected images (1985) and Landsat TM/ETM+/OLI surface reflectance products (1986–2015) were downloaded from the United States Geological Survey’s remote sensing image database (http://earthexplorer.usgs.gov/).

Climate, hydrological and anthropogenic variables were collected and used to quantify the drivers of surface water extent variation. Annual precipitation observed at 50 precipitation stations (Figure 1) was acquired from the Hydrological Year Books of the HRB maintained by Haihe River Water Conservancy Commission and the National Meteorological Information Center of China. Annual streamflow observed at the Gudingqiao hydrological station located upstream to the Cetian Reservoir, Shixiali and Xiangshuipu stations upstream to the Guanting Reservoir were also obtained from the Hydrological Year Books. Anthropogenic variables (

Table 1. Anthropogenic variables used in this study.

Anthropogenic Variables	Unit	Year	Source
GDP	Chinse Yuan (CNY)	1985, 1990–2016	Regional statistical yearbooks
Population	Person	1985, 1990–2016
Electricity energy production	Kilowatt hour	1985, 1990, 1995–2016
Raw coal production	Ton	1985, 1990, 1995–2016
Steel production	Ton	1985, 1990, 1995–2016
Crude iron production	Ton	1985, 1990, 1995–2016
Value of agriculture output	CNY	1985–2016
Urban	km2	1992–2015	CCI-LC data

) were collected from the regional statistical yearbooks, including annual population, gross domestic production (GDP), electricity production, raw coal production, steel and crude iron production, and total value of agriculture output from Datong and Zhangjiakou. Urban areas were calculated from the Climate Change Initiative Land Cover (CCI-LC) dataset that provide land cover maps during long time periods (1992–2015) (http://maps.elie.ucl.ac.be/CCI/viewer/download.php). In order to justify the use of this dataset, the CCI-LC urban cover was evaluated by comparing with urban areas defined by 30-m GlobeLand30 2000 and 2010 datasets, which was reported to have high accuracies [19,20]. Comparison results showed that the difference between the two datasets are within 12%, indicating the feasibility of CCI-LC data.

3.2. Methods

Figure 2 summarizes the workflow of this study. First, annual reservoir water extents were extracted based on Landsat-series imagery using the combination of AWEI/NDWI and Otsu threshold algorithms. In order to characterize the dynamics of reservoir water extents, trend and change-point analyses were then conducted on the time series water extents; variations in the hydrological connectivity between upstream and downstream reservoirs were then analyzed. Third, the contributions of climate variabilities and human activities on the variations of reservoir water extent were quantified using Double Mass Curve (DMC) method. Finally, natural and anthropogenic driving factors of reservoir water extent dynamics were analyzed using correlation analysis. The influence of water management policies on surface water extents were also examined. The software used to implement the procedures is listed in Figure 2.

3.2.1. Image Preprocessing and Extraction of Reservoir Water Extent

For Landsat TM, ETM+ and Landsat OLI, surface reflectance data products were generated using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) and Landsat 8 Surface Reflectance Code (LASRC), respectively [21], which are both atmospheric correction procedures officially developed by NASA Goddard Space Flight Center and the University of Maryland. These products can therefore be directly used for water body extraction. Because there were no surface reflectance products from Landsat MSS, we implemented the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) algorithm embedded in the Environment for Visualizing Images (ENVI) software to correct for atmospheric effect [22]. The data gaps in the Landsat ETM+ images caused by the Scan Line Corrector failure issue were also corrected using Landsat Gapfill algorithm embedded in ENVI software. For each image, the quality assessment band was examined to remove clouds/cloud shadows.

Previous inter-comparison studies suggest that AWEI performs well in both urban area and mountainous areas, and the optimal threshold value is fairly stable across images [13,14]. It has been successfully applied in many recent studies to extract water bodies from Landsat TM/ETM+ images [23,24,25]. Therefore, in this study, we adopted the AWEI approach to extract reservoir water extent from Landsat TM/ETM+ images within the YDRB. Because it can be challenging to separate water pixels from dark surfaces due to their similar reflectance, the AWEI approach provides two indices, namely AWEI_sh and AWEI_nsh. AWEI_sh is suitable for scenes with shadows and/or other dark surfaces, typically found in mountainous regions, while AWEI_nsh is suitable for scenes without shadows. As none of the reservoirs in our study area are located in mountainous regions, only AWEI_nsh is calculated from the Landsat TM/ETM+/OLI images in this study (Equation (1)) [13].

$${\mathrm{AWEI}}_{nsh}=4\times ({\rho }_{Green}-{\rho }_{SWIR1})-\left(0.25\times {\rho }_{NIR}+2.75\times {\rho }_{SWIR2}\right)$$

(1)

where ${\rho }_{Green}$, ${\rho }_{NIR}$, ${\rho }_{SWIR1}$, and ${\rho }_{SWIR2}$ are the surface reflectance of green, near infrared (NIR), shortwave infrared 1 (SWIR1) and shortwave infrared 2 (SWIR2) bands, respectively.

For the Landsat MSS images, NDWI was used instead because they only consist of four spectral bands, namely green, red, and two NIR bands (700–800 um, 800–1100 um). The NDWI is defined as [10]:

$$\mathrm{NDWI}=\left({\rho }_{Green}-{\rho }_{NIR2}\right)/\left({\rho }_{Green}+{\rho }_{NIR2}\right)$$

(2)

After the water index map was created for each image, a threshold value is applied to convert it into a binary map, with “1” denoting water and “0” denoting non-water. In this study, we adopted the Otsu threshold algorithm [26], which is a nonparametric and unsupervised approach for automatic threshold selection. This method determines optimal threshold as the one that maximizes the inter-class variance based on the AWEI or NDWI histogram [27,28]. In our study, the optimal thresholds for the AWEI and NDWI maps ranged from 0.09 to 0.13. Finally, reservoir water extent was extracted over the subset of the binary maps around each reservoir.

The water extent results were then evaluated using high spatial resolution images from Google Earth and Sentinel-2 MSI images with 10-m resolution. Google Earth images were carefully selected so that the acquisition dates were within 10 days from the Landsat images. In the study area, two scenes of 2-m QuickBird images covering Guanting Reservoir (2005) and Yiliuhe Reservoir (2015) were selected as reference image. Sentinel-2 images covering Guanting and Cetian reservoirs (2016) with acquisition dates close to Landsat scenes were also used as reference images. Two thousand sampling points were randomly generated within the 5-km buffer area around each reservoir, visually interpreted as water or non-water, and used as reference datasets.

3.2.2. Trend and Change-point Analyses of Reservoir Water Extents

Once the annual reservoir water extents were extracted, the trend-free pre-whitening (TFPW) Mann-Kendall (M-K) test and Pettit mutation test were adopted to determine the trend and change point of water extents from 1985 to 2016. The Mann-Kendall (M-K) trend test is a nonparametric test method that has been extensively used for time-series hydrological datasets such as runoff [29]. The samples for the M-K test do not need to obey a prescribed distribution and is not sensitive to outliers. In order to eliminate autocorrelation and improve accuracy of trend detection, the TFPW strategy [30] was applied prior to the M-K test. Meanwhile, the Theil-Sen method was used to determine the magnitude of trends in reservoir water surface area and precipitation [31,32].

Change point detection is an important tool for determining if and when a change in a time-series data set has occurred, and has proven to be useful in analyzing the effects of natural factors and human activities on hydrological time series. The Pettitt test [33] is a rank-based nonparametric statistical test method that poses no assumption on the distribution of the data. It adapts the Mann-Whitney test that allows the identification of the time at which the change occurs. Given a time series represented by ($x_1,x_2,\dots ,x_t,x_{t+1},\dots ,x_T)$, which represents the time-series reservoir water extents in this study, the statistical index U_(t,T) is defined as follows [33]:

$$U_{\mathrm{t},T}={\displaystyle \sum _{i=1}^t{\displaystyle \sum _{j=t+1}^T\mathrm{sgn}(x_j-x_i)}},\begin{array}{cc}& \begin{array}{cc}& 1\le t<T\end{array}\end{array}$$

(3)

where

$$\mathrm{sgn}(\theta )=\{\begin{array}{c}1\\ 0\\ -1\end{array}\begin{array}{ccc}& & \begin{array}{c}\theta >0\\ \theta =0\\ \theta <0\end{array}\end{array},$$

(4)

When the time series follows a continuous distribution, the test statistics U_(t,T) can also be calculated by the following recursive iteration:

$$U_{\mathrm{t},T}=U_{\mathrm{t}-1,T}+V_{\mathrm{t},T}\begin{array}{ccc},& \begin{array}{cc}& \end{array}& 2\le t\le T\end{array},$$

(5)

$$V_{t,T}={\displaystyle \sum _{j=1}^T\mathrm{sgn}(x_j-x_t)}\begin{array}{ccc},& \begin{array}{cc}& \end{array}& 1\le \end{array}\mathrm{t}\le T,$$

(6)

The statistics of the most probable change point τ and the significance probability of K_τ are defined as follows:

$$K_{\tau }=\left|U_{\tau ,T}\right|=\max \left|U_{\mathrm{t},T}\right|,$$

(7)

$$\mathrm{p}=2\exp \left(\frac{-6K_{\tau }{}^2}{T^2+T^3}\right)$$

(8)

When p is less than a certain significance level α, we consider that x_τ is a significant change point at level α.

3.2.3. Quantifying Relative Contributions of Climate Variability and Human Activities on Reservoir Water Extents

Climate variability and human interventions are commonly considered two major contributors to variations of water resources in a river basin. In this study, the double mass curve (DMC) method, which was frequently used in hydrological studies, was adopted to quantify their relative contributions to reservoir water extents dynamics. The DMC is a simple visual method composed of cumulative values of two datasets plotted against one another during a given time period [34]. It has been used to evaluate the contributions of potential drivers to changes in streamflow or sediment load [35,36]. Generally, the plot of cumulative streamflow against cumulative precipitation is a straight line if the effect of human activities on streamflow is constant or negligible. Significant changes in DMC slopes indicates departure from the original relationship, suggesting changes in contributions of human activities.

In this study, we plotted the cumulative reservoir water extent (A) against the cumulative upstream precipitation (P). In order to evaluate the change point detected using methods described in Section 3.2, the nonparametric Mann-Whitney U test was used to compare DMC slopes before and after each change point on the curve statistically [37]. Once the significance of the change point is confirmed, regression equations were established based on the data points in the pre-change (reference) and post-change (disturbance) periods. The baseline regression equation in the reference period was then employed to predict the cumulative reservoir water extents for the disturbance period. The difference between the observed and predicted cumulative water extents in the disturbance period was considered as the deviation caused by intensified human activities (∆A_h). Thus, the deviation caused by climate variability (∆A_c) can be calculated as [34]:

∆A_c = ∆A − ∆A_h

(9)

where ∆A, ∆A_c, and ∆A_h are the deviations of annual water extent between disturbance and reference periods, annual water extent deviation caused by climate change, and annual water extent deviation caused by human activities, respectively. The relative contributions of climate variability and human activities to water extents can be determined as:

$$R_c=\frac{\left|\Delta A_c\right|}{\left|\Delta A_c\right|+\left|\Delta A_h\right|}\times 100\%$$

(10)

$$R_h=\frac{\left|\Delta A_h\right|}{\left|\Delta A_c\right|+\left|\Delta A_h\right|}\times 100\%$$

(11)

where $R_c$ and $R_h$ are the relative contributions of climate variability and human activities to water extents, respectively.

3.2.4. Correlation Analysis of Driving Factors and Reservoir Water Extents

For both pre-change and post-change periods, Pearson correlation coefficients were calculated to examine the effect of annual precipitation and socio-economic factors on reservoir water extent dynamics. Here precipitation was treated as a natural driving factor. Human factors include GDP, population, electricity production, raw coal, steel and crude iron productions, total value of agriculture output and urban area, as they are all associated with water usage. For the Guanting reservoir, the values of human factors were calculated as the sum of Datong and Zhangjiakou, as they are the two major upstream cities. For the Cetian Reservoir, the values for Datong was used. Steel and crude iron production was not included for the Cetian reservoir because of its low production (average annual production < 40 × 10⁴ ton). High correlation coefficients indicate stronger possible impacts on variations in reservoir water extent. The differences between pre- and post- change coefficients suggest that the main driving factors of the reservoir area might have changed. The correlation coefficient between upstream and downstream reservoir water extents can be used to assess the hydrological connectivity in the river basin. Higher correlation coefficients indicate stronger connections, and can thus be interpreted as stronger influence of the upstream on downstream.

4. Results

4.1. Spatiotemporal Dynamics of Surface Water Extents in Reservoirs

The water extents of the 14 reservoirs in YDRB were extracted based on the 107 Landsat-series images from 1985 to 2016 using AWEI or NDWI approaches. Using Google Earth images as references, the overall accuracies vary from 98.8% to 99.7%, and the kappa coefficients vary from 0.95 to 0.98 (

Table 2. Accuracies of the extracted reservoir water extents.

Reservoir	Year	Reference Image	PA (%)	UA (%)	OA (%)	Kappa Coefficient
Guanting	2005	QuickBird (2 m)	93.4	97.2	98.8	0.95
	2016	Sentinel-2 (10 m)	94.6	96.8	98.9	0.95
Yiliuhe	2015	QuickBird (2 m)	98.0	97.3	99.7	0.97
Cetian	2016	Sentinel-2 (10 m)	98.1	98.7	99.8	0.98

PA: producer’s accuracy; UA: user’s accuracy; OA: overall accuracy.

).

Figure 3 shows the frequency of surface water occurrence in the study area, calculated as the number of years that a pixel has been detected as open water, divided by the total number of years (i.e., 32 years in this study). The statistics of annual surface water extents in the 14 reservoirs are shown in Figure 4. The Guanting and Cetian reservoirs are the two biggest reservoirs in the YDRB. The average water extents in the Guanting and Cetian reservoirs are 68.6 km² and 16.8 km², respectively, which account for over 75% of the total water extent in the basin. Both reservoirs show great variations in water extents, ranging from 31.9 km² to 121.3 km² in Guanting and 6.5 km² to 33.8 km² in Cetian. Other reservoirs are very small, with an average area of less than 5 km², and several of them less than 1 km². Nonetheless, almost all reservoirs showed similar temporal dynamics during the whole study period, with increase in their water extent at the beginning and decrease afterwards. Most reservoirs reached their maximum areas in 1995 or 1996, including Guanting, Cetian and Youyi.

Figure 5 showed inter-annual variability in water extents in both Guanting and Cetian reservoirs (see Figure A2 and Figure A3 for the maps of annual water extent). The water extent dynamics of Guanting can be divided into four stages: slow rising period, severe atrophy period, stable period and recovery period. In the slow rising period of 1985–1995, its area increased from 66.7 km² to 121.3 km², with an expansion of 81.9%. In the severe atrophy period of 1996–2007, its area decreased sharply from 113.2 km² to 31.9 km², shrinking by 71.8%. From 2008–2012, the area was relatively stable, with a mean area of 36.2 km². During 2013–2016, the reservoir area rebounded slightly, increasing from 41.6 km² to 77.1 km², with an average area of 62.9 km². The area of Cetian is much smaller than the Guanting reservoir, but varied significantly over the study period figuring three states (Figure 4): a slow rising period (1985–1995), a shrinking period (1996–2011) and a recovery period (2012–2016). During the slow rising period of 1985–1995, the water extent increased from 11.8 km² to 33.8 km², with an average annual increase rate of 2.2 km²/year. In the shrinking period from 1996 to 2011, the reservoir area decreased from 33.3 km² to 9.3 km², reducing by 72.1%. During the 2012–2016 period, its area recovered from 10.2 km² to 15.6 km². Given that the total surface area of all other reservoirs is too small, Guanting and Cetian were selected as reservoirs for analyzing variability of surface water dynamics in this basin.

4.2. Trend and Change-points of Dynamic Water Extents

The TFPW-MK trend test showed that the water extents in Guanting and Cetian had a significant downward trend from 1985 to 2016 (

Table 3. Trend and change-point of reservoir water extents and streamflow from 1985 to 2016.

Reservoir		TFPW-MK Trend Test		Pettit Change-Point Test
		Slope (km2/year)	Z	Year	P
Guangting	Water extent (km2)	−1.56	−4.86 **	2001	0.00 **
	Streamflow (m3/s)	−0.28	−4.21 **	1999	0.00 **
Cetian	Water extent (km2)	−0.116	−2.24 *	2005	0.03 *
	Streamflow (m3/s)	−0.067	−2.45 *	2005	0.01 **

* Significant at the 0.05 level; ** Significant at the 0.01 level.

), and the annual average declination rate were 1.56 km²/year for Guanting and 0.116 km²/year for Cetian, respectively. The Pettitt test showed that the significant change-point of time-series water extent occurred in 2001 (p < 0. 01) for Guanting and 2005 (p < 0.05) for Cetian (Table 3). For the Guanting Reservoir, the average water extent was 88.5 km² in the pre-change period of 1985 to 2001 and 46.1 km² in the post-change period of 2002 to 2016, decreasing by 47.9%. The average water extent of Cetian was 18.8 km² in the pre-change period of 1985 to 2005 and reduced to 12.9 km² during the post-change period of 2006 to 2016, decreasing by 31.4%. Interestingly, from Figure 5 we can visually detect an apparent upward trend in Guanting since 2012. Nonetheless, the Pettit test on the water extents during the post-change period (2002–2016) indicated 2012 as the most probable change-point with p-value of 0.18 due to the small sample size.

As illustrated in Figure 6a,b, both Guanting and Cetian reservoirs demonstrated similar temporal trends in their annual inflow observed from their upstream hydrological stations. Note that the Guanting inflow was calculated as the sum of streamflow observed at the Shixiali and Xiangshuipu hydrological stations (Figure 1). The TFPW-MK trend test showed that the streamflow also demonstrated significant downward trend with change points occurring in 1999 for the inflow of Guanting and 2005 for the inflow of Cetian. For the Cetian Reservoir, the correlation analysis showed a significantly positive relationship between annual streamflow and reservoir water extent. The Pearson correlation coefficients are 0.81 (p < 0.01) for the whole study period from 1985 to 2016, 0.77 (p < 0.01) for the pre-change period (1985–2005), and 0.72 (p < 0.05) for the post-change period (2006–2016). For the Guanting reservoir, the positive correlations are statistically significant for the whole study period (r = 0.83, p < 0.01) and pre-change period (r = 0.78, p < 0.01) (1985–2001), while the correlation coefficient for the post-change period (2002–2016) is 0.49 and p-value is 0.07. Correlations between water extents in Guanting and Cetian extents are 0.55 (p < 0.01) for the 32 years, and 0.72 (p < 0.01) during the Guanting pre-change period (1985–2001) and 0.25 (p > 0.05) during the post-change period (2002–2016).

4.3. Relative Contributions of Climate Variability and Human Activity to Reservoir Water Extents

As shown in Figure 7, the linear regression lines between the cumulative water extents and the cumulative annual precipitation have different slopes before and after the change point. By performing Mann-Whitney U tests [36], the slopes are confirmed to be statistically different, and the change points detected by the Pettit test are reasonable (p < 0.01 for both reservoirs).

Water extents in Guanting and Cetian reservoirs increased 5.8 ± 16.4 km² and 0.8 ± 3.1 km², respectively, in response to climate variability. The relative contributions to the changes in water extents being 25.9 ± 22.4% and 24.9 ± 17.3%. Human intervention led to decreases in annual water extents in Guanting and Cetian reservoirs of 48.1 ± 20.0 km² and 6.7 ± 3.0 km², respectively, contributing 74.1 ± 22.4% and 75.1 ± 17.3% to the changes in reservoir water extents (

Table 4. Temporal variation of the relative contributions of climate variability and human interventions to annual water extents in Guanting and Cetian reservoirs.

Reservoir	Periods	ΔA (km2)	ΔAc (km2)	ΔAh (km2)	ΔA/A (%)	Rc (%)	Rh (%)
Guanting	2002–2016	−42.4	5.8 ± 16.4	−48.1 ± 20.0	−47.8	25.9 ± 22.4	74.1 ± 22.4
	2002–2012	−48.5	0.9 ± 15.3	−49.3 ± 20.5	−54.8	25.3 ± 25.3	74.7 ± 25.3
	2013–2016	−25.6	19.3 ± 12.0	−44.9 ± 21.0	−28.9	27.6 ± 13.6	72.4 ± 13.6
Cetian	2006–2016	−5.9	0.8 ± 3.1	−6.7 ± 3.0	−31.6	24.9 ± 17.3	75.1 ± 17.3
	2006–2012	−6.8	−0.6 ± 2.3	−6.2 ± 2.9	−30.0	25.5 ± 20.0	74.5 ± 20.0
	2013–2016	−4.3	3.2 ± 3.0	−7.5 ± 3.4	−22.6	23.8 ± 14.3	76.2 ± 14.3

Note: ΔA is the difference in water extent between the reference and disturbance periods. ΔA_c and ΔA_h are the water extent changes attributed to climate variability and human activities, respectively. Rc and Rh are the relative contributions of climate variability and human activity to reservoir water extents.

). Evidently, contributions of human activities were much higher than those from climate variability in both reservoirs after the change point. For Guanting, the disturbance period can be further divided into two sub-periods, i.e., 2002–2012 and 2013–2016, based on the distinct expansion rates of water extents (Figure 5). The negative contribution of human activities during 2002–2012 was higher than that during 2013–2016. Our results suggest that human activities and climate variability played opposite roles in variations of water extents in both reservoirs, while the negative effect of human activities seems to decrease after 2012.

4.4. Correlations between Driving Factors and Reservoir Water Extents

The graphical trends of human factors and water extents (Figure 8 and Figure 9) showed that GDP, population, urban area, electricity energy production and value of agriculture output in the upstream regions showed a steady positive trend since 1985, with a rapid increase in the slope after 2000 for both Guanting and Cetian reservoirs. The raw coal production peaked in 1995, decreased from 1995 to 2000, followed by a sharp increase since 2001, matching the expansion and shrinkage of the water extents in both reservoirs (Figure 8e and Figure 9e). For the Guanting reservoir, the upstream regions exhibited sharp increase in steel and crude iron production since 2000, which becomes stable after 2012.

Table 5. Pearson correlation coefficients between water extents and precipitation & human activity factors in Guanting reservoir and Cetian reservoir.

	Guanting Reservoir			Cetian Reservoir
	1985–2001 (Pre-Change)	2002–2016 (Post-Change)	2002–2012	1985–2005 (Pre-Change)	2006–2016 (Post-Change)	2006–2012
Precipitation	0.680 **	0.413	0.229	0.610 **	0.308	−0.194
GDP	0.246	0.336	−0.564	0.264	0.162	−0.884 **
Population	0.201	0.300	−0.519	0.202	0.136	−0.755 *
Electricity energy production	0.373	0.246	−0.516	0.373	0.004	−0.786 *
Raw coal production	−0.492	0.258	−0.304	0.492 *	0.126	−0.771 *
Steel production	0.020	−0.005	−0.833 **	--	--	--
Crude iron production	0.334	0.082	−0.722 **	--	--	--
Value of agriculture output	0.538 *	0.43	−0.502	0.536	0.269	−0.740
Urban area	0.217	0.350	−0.651 *	0.217	0.211	−0.851 *

* Correlation is significant at the 0.05 level; ** Correlation is significant at the 0.01 level.

lists the Pearson’s correlation coefficients between water extents and the drivers for both reservoirs. Precipitation had significantly positive relationships (p < 0.01) with water extents during the pre-change periods for both reservoirs, while after the break points the positive correlations become insignificant. It is interesting that none of the factors had significant correlations with water extents during the post-change periods. As Figure 6 indicates evident water extent expansions after 2012, we further analyzed the correlations between the break points and 2012. It can be seen that GDP, population, electricity energy production, raw coal, steel and crude iron productions, value of agriculture output and urban area are all negatively correlated to water extents in Guanting reservoir during this period. The scatterplots of the water extents against the driving factors (Figure A4 and Figure A5) showed negative slope of the linear regression line. In Cetian, the negative correlation coefficients are statistically significant (p < 0.05), and the regression line has R² greater than 0.55, suggesting negative effects of human activities on water extents.

5. Discussions

5.1. Characteristics of Reservoir Water Extents Dynamics

5.1.1. Trend of Reservoir Water Extents and Their Representativeness of Water Resources in YDRB

The Guanting and Cetian reservoirs are the two biggest surface water bodies in the YDRB. From 1985 to 2016, the water extents in both reservoirs experience similar temporal patterns. They both expanded from 1985 to 1995, reached the maximum extent in 1995, and then shrank until 2011. From 2012, the water extents of both reservoirs started to rebound. The combination of the TFPW M-K and Pettit tests showed that both reservoirs experienced downward trend over the 32 years, and the abrupt change points occurred in 2000s. Our results for Guanting reservoir are consistent with a recent study, which reported that the water level in its hydro-fluctuation zone had dropped 8.19 m from 1996 to 2007 [38].

For both reservoirs, the time-series of water extents showed consistent temporal patterns with the streamflow observed from upstream hydrological gauging stations. The positive correlation between streamflow and water extents were statistically significant throughout the study period (p < 0.01). Trend analyses on annual streamflow also showed that abrupt changes occurred closely with the reservoir water extents. This consistency suggests that surface water extents can be used to represent variations in water resources in YDRB.

Previous studies have suggested that surface water extents can be used as a proxy for streamflow in hydrological modelling. Revilla-Romero et al. (2016) [39] assimilated satellite-derived surface water extents into a rainfall-runoff model to update simulated streamflow, and demonstrated that such an integration improved streamflow simulations. Similarly, Revilla-Romero et al. (2015) [40] employed remotely-sensed surface water extents from the Global Flood Detection System (GFDS) to calibrate a hydrological model and demonstrated the potential of the satellite-derived water extent for improving hydrologic model calibration in ungauged or poorly gauged watersheds. Huang et al. (2014) [41] reported high correlation coefficients between satellite-derived wetland water extents and the stream flow, and suggested that inundation maps can be used for improving model studies of drought and stream runoff. The findings of these studies, as well as our results, suggest that the satellite-derived water extents can be considered as feasible indicators for time-series water resource dynamics in YDRB.

5.1.2. Hydrological Connectivity Discovered from Reservoir Water Extents Dynamics

Correlation analyses between upstream and downstream streamflow observed at gauge stations have been frequently used to examine hydrological connectivity in a watershed [4,42]. The correlation coefficient of reservoir water extents between Cetian and Guanting reservoirs was statistically significant over the past three decades (p < 0.01), indicating strong connections between upstream and downstream reservoirs in YDRB. That is, variability in upstream reservoirs has a strong influence on downstream reservoirs.

However, we found that the correlation in the pre-change period (r = 0.715, p < 0.01) was stronger than that in the post-change period (r = 0.254, p > 0.05). Although showing similar temporal trends, the downstream Guanting reservoir experienced more rapid shrinkage than the upstream Cetian reservoir. In addition, the abrupt change in Guanting water extent occurred four years earlier than that of Cetian. This indicates that the faster drying of Guanting can be attributed to more intensified upstream human activities.

Jiang et al. (2014) [4] studied the hydrological connections in the Beijing section of Yongding River, which is located in the lower reach starting from Guanting. They reported that the hydrological connections between Guanting and the downstream Sanjiadian had declined since 2000 due to the prolonged zero flows below Sanjiadian. In our study, we found that the hydrological connections above the Guanting Reservoir also declined during the post-change period, which is likely to contribute to the zero flows in the lower reach of YDRB.

5.2. Relative Contributions of Climate Variability and Human Activities to Reservoir Water Extents

Our study showed that the relative contributions of human activities to water extents in both reservoirs were around 74%~75% compared to those in the pre-change period for both reservoirs. Previous studies have investigated the contributions of climate variability and human activities to runoff variations in HRB based on the streamflow observations at gauge stations. Liu et al. (2017) [5] and Xu et al. (2014) [43] reported that 66.2%~73.1% of runoff reduction in HRB was attributed to human activities based on the Budyko hypothesis model. Wang et al. (2013) [44] evaluated four catchments in HRB, and reported that human activities contributed to 63.1% of runoff reduction. Xia et al. (2014) [45] studied the runoff variations in YDRB from 1960 to 2010, and reported that the contributions from human activities to runoff declination accounted for 71.0%~92.7%. Although our study focused on the reservoir water extents instead of runoff, we have shown that the reservoir water extents correlated significantly with runoff, and the relative contributions of climate variability and human activities to water extents were similar. Most previous studies have focused on periods before 2010. Our analysis updates the knowledge on changes in water resources and their driving factors in YDRB, highlighting the prolonged intensification of human activities in the recent decades.

5.3. Drivers for Reservoir Water Extent Dynamics

Precipitation had significant positive influences on the reservoir water extents variations during the pre-change period. During the post-change period, the positive influence become weaker, confirming that precipitation had a minor influence on the declination of reservoir water extents. From the change point to the year 2012, almost all socioeconomic factors were negatively associated with water extents. Since 2000, YDRB has witnessed increasingly rapid socioeconomic development. For the upstream regions of Guanting Reservoir, the growth rate of GDP was 2.4 billion Yuan RMB per year from 1985 to 2001, which increased to 14.3 billion CNY per year from 2002 to 2016. During the post-change period, the growth rate of GDP in the upstream region of Cetian reservoir increased more than twice than that of the pre-change period. Population and urban area also increased more rapidly after the change point. The fast economic development, population growth and urban expansion led to increase of domestic and industrial water demands in the basin. For both reservoirs, human-induced water extents changes were also negatively correlated with crop production. Increasing crop production increases water usage for agriculture irrigation and hence decreases reservoir water storage. In YDRB, the primary industrial water consumption comes from coal power generation, metallurgy, and food production [46]. Steel and crude iron are the mainstay industry in Zhangjiakou city in Hebei Province, which is located in the upper reaches of the Yonding River. Datong city in Shanxi Province, which is close to the Cetian reservoir, is home to one of the largest coal deposits in China. Coal mining has been the mainstay economic industry in Shanxi Province, and is also a major surface water consumer [47]. Therefore, the sharp increase in productivity of mainstay industries in the upstream regions, i.e., raw coal for Cetian and steel and crude iron for Guanting led to water extents declination in the reservoirs.

5.4. Effects of Policies on Reservoir Water Extents

Table 5 showed that human activities correlated with reservoir water extents negatively during the period from the change-point year to 2012, while their effects were not significant in the entire post-change period. The water extents in both reservoirs have recovered since 2012, despite continuous rise of socioeconomic indicators. While this can be partially attributed to increasing precipitation from 2012, the implementation of water-related policies also contributed to the recovery of water extents.

In 2011, the Shanxi Wanjiazhai Yellow River Diversion Project in the northwest region of Shanxi Province was completed [48]. It was one of the strategic measures by the ShanXi Province to transfer water from the Yellow River and alleviate water shortages in the three industrial cities of Datong, Taiyuan, and Pingsuo. The project brought 160 million cubic meters of water to Datong each year. The supply of the Yellow River water reduced water deficit in Datong, therefore allowing more water to drain downstream, leading to the rebound of water extent in the Cetain reservoir. Meanwhile, the middle route of the South-to-North Water Diversion Project, which originates from Hanjiang river, began its operation in December 2014. By the end of 2016, the volume of water coming to Beijing from the South Water has reached 2.3 billion cubic meters, of which 1.7 billion cubic meters was supplied to tap water plants, and the remainder directed to reservoirs, emergency water sources, rivers and lakes (http://www.bjwater.gov.cn). All these policies help alleviate the pressure of water demand from the reservoirs; hence, the surface areas of Guanting and Cetian were able to recover.

These policies affected not only the inter-annual but also intra-annual variability of reservoir water extents. After 1996, the storage of the Guanting reservoir continued to decrease. In order to rationally allocate limited water resources in the YDRB, in 2001, the Government of Beijing and the Ministry of Water Resources of the People’s Republic of China authorized the "Capital Water Resources Sustainable Utilization Plan in the Early 21st Century" [49]. To achieve this goal, from 2003, the Cetian reservoir began to transfer water to the Guanting reservoir (

Table 6. The outflow and dates of water delivery from Cetian reservoir to Guanting reservoir.

Year	Cetian Outflow (×104 m3)	Date	Year	Cetian Outflow (×104 m3)	Date
2003	5010	09/26–10/07	2010	1500	10/15–11/16
2004	7140	10/12–11/15	2011	--	--
2005	6703	10/17–11/06	2012	--	--
2006	1579	10/13–11/22	2013	3000	10/08–11/11
2007	2601	10/10–10/22	2014	3000	11/07–11/22
2008	1800	10/10–10/20	2015	2018	09/29–11/23
2009	--	--	2016	2200	10/17–10/30

). Except for 2009, 2011 and 2012, water was delivered during early-October to mid-November each year. The total amount of water delivered reached 393 million m³ since 2003. The water delivery caused distinct variations in reservoir water extents during the two months. Figure 10 demonstrates the water extent before and after water delivery from 2002 to 2008 observed from Landsat-series imagery following the same procedure in Section 3.2.1. Starting from 2003, the area of Guanting increased obviously after water delivery, and the water extent expansion was proportional to the amount of water inflows. The area of Cetian decreased after water delivery, and the water extent shrinkage was also proportional to the water output. As a reference, the water extents in both reservoirs did not change during September to November in 2002, when centralized water delivery was not initiated.

6. Conclusions

In this study, the dynamics of surface water extents in reservoirs of YDRB was characterized based on time-series of Landsat 4 MSS, Landsat 4-5 TM, Landsat 7 ETM+ and Landsat 8 OLI images during 1985-2016 using a combination of AWEI, NDWI and Otsu threshold algorithms. All reservoirs demonstrated considerable variations in water extents over the past 32 years. The downstream Guanting and upstream Cetian reservoirs were identified as the biggest open water bodies, with a combined water area of over 75% of the total reservoir areas in the basin.

We proved that the reservoir water extents in YDRB are representative of variations in surface water availability, as the annual inflow to each reservoir and the reservoir water extents showed consistent temporal trend and had significant positive correlation. The time series approaches, namely TFPW M-K test and Pettit test, which were previously used for streamflow analysis, were then applied to examine the temporal trend and change-point of water extents dynamics. The DMCs, which were previously used to assess the trend of river discharges, were further adopted to analyze the relative contribution of human activities and climate change on water extent variability. Results showed that both reservoirs experienced statistically significant downward trend in water extent, and the most probable change points occurred in 2001 and 2005 for the Guanting reservoir and Cetian reservoir, respectively. The relative contribution of human activities to shrinkage of water extents was 74%~75% for both reservoirs during the post-change period.

Precipitation, socio-economic factors and policies are all drivers for variations in reservoir water extents, while they played different roles in different periods. Precipitation is the major driving factor during the pre-change periods. From the change-point year (2006) to 2012, human drivers including GDP, population, electricity power production, raw coal production, value of agriculture output, and urban area played more important roles and negatively influenced water extents in Cetian. For the downstream Guanting reservoir, steel and crude iron production was also an important factor.

From the remotely-sensed water extents, we identified a reduced hydrological connectivity between the upstream Cetian reservoir and downstream Guanting reservoir during the post-change period. The effects of inter-basin and intra-basin water transfer projects implemented by regional governments, including the South-to-North Water Diversion Project, were also revealed by the recovery of water extents since 2012 and the intra-annual change of the water extents.

Existing studies in the remote-sensing community mainly focused on methodological development for water extent extraction, while studies in hydrology communities usually relied on streamflow measurements at hydrological stations to analyze the dynamics and driving factors of water resources. The framework presented in this study linked the remote-sensing techniques and hydrological time series analysis approaches to assess the dynamics and driving factors of water resource variations. Compared to streamflow observations in hydrological stations, remote-sensing technique features high efficiency and low cost. This is potentially useful in un-gauged or poorly-gauged watersheds. The hydrological time series approaches help quantify the contributions of natural and anthropogenic factors to variations of water availability in the basin, which could provide guidance for better decision-making for water resources management in the basin.