The monthly grids from GRACE Tellus land data are applied to analyze TWS variations. The product is derived from the latest Release-05 spherical harmonics, which is an improvement over the previous 04 version. Several institutions provide gravity solutions, such as the University of Texas Center for Space Research (CSR), NASA's Jet Propulsion Laboratory (JPL), and Deutsches GeoForschungsZentrum (GFZ). A recent comparison suggested that TWS estimates from GFZ, CSR, and JPL solutions were highly correlated with one other, and tiny differences among them were within the margin of solution's error (Sakumura et al., 2014). Among these three products, the one from CSR had the smallest root-mean-square (RMS) of deviations between the ensemble mean and itself in 156 basins around the world. In this study, we chose products derived from CSR's solution for the following analyses. The paper by Swenson and Wahr (2006) describes the details of the post-processing for the spherical harmonics. The final grid (1∘ in both latitude and longitude) values are presented in the form of changes in equivalent water thickness (unit: cm) relative to a time–mean baseline. The data period is from January 2003 to December 2013 and months of absent data are as follows: June 2003, January and June 2011, May and October 2012, and March, August, and September 2013. Grid-scale factors, which correspond to the gridded product, were used to partially correct leakage errors and restore the amplitude-damping caused by the filtering process. Errors and uncertainties in mass variation can be computed from the scaled gridded data (Launderer and Swenson, 2012). In this paper, gridded fields of scale factors and error estimates provided along with the GRACE Tellus products were calculated following Landerer's method based on NCAR's CLM4 model (Oleson et al., 2008). Recently, Long et al. (2015) conducted comprehensive comparisons to assess skills and uncertainties of different approaches for processing GRACE data to restore signal losses caused by spatial filtering.

Monthly flux/state variables (Table 2) from GLDAS (Rodell et al., 2004) were applied to estimate water storage variations, and these variables were also used to address variations in the components of water storage. GLDAS drives four land surface models: Mosaic (Koster and Suarez, 1996), Noah (Chen et al., 1996; Koren et al., 1999; Betts et al., 1997; Ek et al., 2003), Community Land Model (CLM) (Bonan et al., 1998; Dickinson et al., 1993; Dai and Zeng, 1997), and Variable Infiltration Capacity (VIC) (Liang et al., 1994, 1996). Satellite-based and ground-based observations are integrated into these models to generate optimal fields of land surface states and fluxes. The forcings for these models from 2001 to present are a combination of NOAA/GDAS atmospheric analysis fields, spatially and temporally disaggregated NOAA Climate Prediction Center Merged Analysis of Precipitation (CMAP) fields (Xie and Arkin, 1996) and observation-based downward shortwave and long-wave radiation fields from the Air Force Weather Agency (AFWA).

Drainage networks are mostly distributed in the monsoon-dominated middle and east China (Fig. 1a), which are also highly populated areas with high levels of water consumption. In this study, we specifically focus on eight large basins: Heilongjiang River, Liao River, Hai River, Huai River, Yellow River, Changjiang River, Zhujiang, southeast rivers (with abbreviations of HLJ, LR, HaiR, HuaiR, YR, CJ, ZJ, and SERs, respectively, in the following tables). Desert is the dominant land cover in northwestern China, while glacier, snow cover, and frozen soil are widely distributed across the Tibetan Plateau (Fig. 1b). Vector data for the desert are acquired from the Data Sharing Infrastructure of Earth System Science (http://www.geodata.cn). The Second Glacier Inventory Dataset of China (Version 1.0) is acquired from Science Data Center for Cold and Arid Regions (http://westdc.westgis.ac.cn). Annual Chinese water resources bulletins from 2003 to 2012 are acquired from the Ministry of Water Resources to assist with the analysis. Values for surface water resources, groundwater resources and gross water resources provided in the water resources bulletins are the results of existing monitoring and statistical analyses (Fig. 5). Surface water resources refer to water storage in rivers, lakes, and glaciers, and groundwater resources mainly refer to water storage in underground shallow aquifers.

The effect of the truncation (Ordermax=60) and filtering processes (300 km Gaussian filtering) on the GRACE spherical harmonics is equivalent to a low-pass filter; thus the effective resolution of the GRACE TWS product is several hundred kilometers (Tapley et al., 2004). The TWS time series in a 1∘ grid was mixed with TWS signals from the surrounding area, leading to leakage errors. When the outside TWS signal was stronger than the inside, the grid value was exaggerated by leakage errors and vice versa. In addition, the sign of gridded TWS could even be changed in cases where inside and outside TWS signals had opposite phases caused by extreme changes in topography, such as in the Turpan basin in northwestern China. As relationships between TWS series at different spatial scales were inferred from land hydrology model simulations, grid-scale factors calculated based on this information could partially correct GRACE TWS and to some extent recover small-scale information (Landerer and Swenson, 2012); thus, these scale factors can be quite helpful for extracting TWS over arbitrary shaped region.

The RMS value of TWS time series in a specific grid is an indicator for the amplitude of local TWS. And the empirical probability density distribution (empirical PDF) curve for RMS values in the research region described the statistical distribution of TWS amplitude within the area. In Fig. 2, empirical PDF curves based on TWS data from modeled TWS data (MOSAIC, VIC, CLM, NOAH, and GLDAS ensemble mean) and observation TWS data (scaled and unscaled GRACE data) were compared. Empirical PDF curves based on scaled GRACE data and modeled data (except CLM) all showed a larger RMS value range on the x axis than that based on unscaled GRACE data. This means the range of TWS amplitude within the research area has been stretched after scaling. In addition, empirical PDF curves based on scaled GRACE data and most modeled data showed that RMS values concentrated in the relatively low numerical zone, with lowest RMS values close to 0 cm. Spatially, areas with low RMS values correspond to northwest China, which is an arid climate zone with vast deserts (Figs. 1, 3b and c). From the comparison in Fig. 3, we can also see that scaled GRACE TWS has a similar distribution of amplitude to that from the GLDAS ensemble mean over China, particularly the boundary with RMS of 3cm, separating arid and humid climate zones. TWS is quite stable over some part of the oceans and major deserts around the world; thus a small RMS for TWS in these areas indicates small data noise in GRACE TWS (Sakumura et al., 2014). Both the empirical PDFs and the spatial distribution of the RMS of the TWS suggested that grid-scale factors could correct the amplitude of TWS in space. Previous research has demonstrated that correction for leakage is critical to regional TWS analysis (Chen et al., 2014).

Regionally averaged TWS time series from scaled and unscaled GRACE data are highly correlated, and this means that the fluctuation process in TWS has not been heavily influenced by the scale factors. At the same time, the values of TWS were all amplified to different degrees (Table 5), with the amplitudes in the Huai River basin and Zhujiang basin increasing over 50 %. However, in the Liao River basin and Yellow River basin, only small changes occurred (< 10 %). The slopes in Table 5 can be regarded as the basin-specific scale factors for GRACE TWS. Generally, basins with large area are less affected by leakage errors and have slopes close to 1.

In general, fluctuations in annual precipitation could appropriately characterize the interannual variability in regionally averaged TWS, but distinct processes also exist in certain basins or over certain periods because of the influence of other factors (Figs. 4, 6). TWS in China was at a relatively high level before 2006, but then stayed at a continuously low level with a high variability from 2006 to 2012. During this period, severe droughts occurred frequently and caused particularly sharp declines in TWS in 2007, 2009, and 2011. TWS in China did not recover to the same level as in 2005 until 2013. This periodic process was partially reflected in the TWS estimates from the GLDAS ensemble mean (mostly soil moisture storage) but not in the residual series or water resources records (Figs. 4, 5).

In northeast China, TWS observations and simulation estimates in the Heilongjiang River basin were consistent and showed no long-term trend; the region mainly suffered from two severe regional droughts in 2007 and 2011 and a significant basin flood event in 2013. Annual precipitation in the Liao River basin continuously declined from 2005 to 2009 and then increased rapidly in following years; TWS estimates and water resources records both captured this process precisely. Nevertheless, it seemed that the TWS observations failed to respond to heavy precipitation in 2010, with a 9 cm increase in TWS estimates and only a 3 cm increase in TWS observations.

In north China, annual precipitation in Hai River basin changed following a V-shaped process, with year 2006 as the turning point, which can also be recognized from the TWS estimates and water resources records. After a rapid decline from 2004 to 2006 (-2.48 cmyr-1), the TWS from the scaled GRACE data in Hai River basin became stable around 2007. Contrary to increasing precipitation, TWS dropped 3 cm in 2008 and continued to decline slowly at a rate of -0.22 cmyr-1 until it started to recover in 2012. The TWS in Hai River basin generally showed a linear decreasing trend (-1.27 cmyr-1) during 2004–2011. The residual between TWS from scaled GRACE data and the GLDAS ensemble mean could be treated approximately as the sum of surface reservoir and groundwater storage. Moreover, detection depth of GRACE is much deeper than the layers considered in models (1.90 m in VIC, 2.00 m in NOAH, 3.50 m in Mosaic, and 3.43 m in CLM) and in field monitoring (shallow aquifer). Although the increasing precipitation seemed to have alleviated the depletion trend in these areas, we should not ignore the large gap (-1.80 cmyr-1) between the trends of the time series of the residuals and summed water resources records (-1.28 and 0.50 cmyr-1) during 2006–2011. The gap between these time series probably suggests that the long-term effect of over-exploitation of groundwater still remained, even though water-saving management practices had already been carried out in this basin, and water storage would suffer even worse depletion in the future drought years. Similar to the Hai River basin, the TWS from scaled GRACE data in the Yellow River basin followed a nearly linear decreasing trend (-0.73 cmyr-1) during 2004–2011, and it changed more slowly (0.13 cmyr-1) after 2007. The basin-averaged TWS, gross water resource, and precipitation also showed different processes in the latter half of research period. However, Fig. 7 reveals that the area with large long-term decreasing trends is mainly located midstream of the Yellow River basin (Shanxi and Shaanxi provinces), which is famous for coal mining. To identify the exact causes for decreasing TWS, more local statistical data and groundwater level records should be collected. Over recent decades, the TWS in the Huai River basin has shown a long-term descending trend (-0.56 cmyr-1), which is similar to annual precipitation over this basin. The TWS from GLDAS ensemble mean also showed good agreement.

Annual variations in TWS from scaled GRACE data, the GLDAS ensemble mean, and water resources records are more similar across basins in south China than they are in north China. The TWS in the Changjiang River basin, Zhujiang basin, and southeast rivers basin all followed an increasing trend from 2003 to 2013, at 0.35, 0.55, and 0.70 cmyr-1, respectively. As a result of typhoons and tropical storms, the TWS in these basins also had much stronger fluctuations than in north China.

The TWS variations observed by GRACE are integrated information from different components in the vertical profile. Compared to water storage in surface soil layers, snow cover and canopy water storage are almost negligible in most regions, and analysis results in China suggested that changes in soil moisture contributed significantly to TWSC (Table 6). The percentage of TWSC variance explained by SMC could be as high as 62 % at the national scale. In most basins, over half of the TWSC variance could be attributed to SMC, with high percentages in the Changjiang River basin and Zhujiang basin (64 and 67 %). In the Heilongjiang River basin, SMC played a less important role in TWSC (38 %). Fluctuations in hydrologic fluxes over the basin jointly affected water storage. According to correlation analysis based on GRACE TWS and GLDAS fluxes, precipitation, evapotranspiration, and runoff each contributed 46, 41, and 32 %, respectively, to the TWSC variance in China (Table 6). As most basins we focused on are under control of the monsoon climate, precipitation and runoff generally showed higher contributions to the TWSC variance than evapotranspiration did. Overall, precipitation was found to have a much higher impact on TWSC in the south than in the north, with the highest explained variance in the Zhujiang basin (60 %), followed by that in the Changjiang River basin (44 %).

When focusing on differences between large regions, spatial patterns of linear trends calculated from scaled and unscaled GRACE TWS are consistent (Fig. 7a, c). But at a local scale, results from scaled GRACE TWS better correspond to natural features of the TWS intensity distribution. Areas around river networks usually have a high level of TWS, thus presenting large absolute values of trends. From 2003 to 2013, four main regions were identified with intensive and significant long-term trends in TWS. Results also revealed that seasons in a year made different contributions to these trends (Fig. 8).

According to the analysis in the previous section, we inferred that human activities rather than climate parameters could be responsible for the significant TWS depletion in north China, as withdrawals usually surpass net recharge in arid and semiarid regions. Severe areas are mainly located in the province of Shanxi and the southern part of the Hebei province, with decreasing trends less than -0.80 cmyr-1. And offset to loss rate caused by mass gains from reservoir regulation, water diversion and coal transport in this region was estimated to 0.76 cmyr-1 (Tang et al., 2013). In the Shanxi province, the east to mid-section of Yellow River, coal mining not only has disturbed normal recharge to the nearby aquifer but has also caused over-exploitation of groundwater. The groundwater is a major source of water consumption in the Huang–Huai–Hai Plain; agricultural irrigation consumes large amounts of freshwater pumped from deep wells every year (Foster et al., 2004; Kendy et al., 2004). This poor condition deteriorates with seasons and the depletion becomes most severe, impacting the largest area in autumn. Considering irrigation demand concentrated mostly in MMA and high social water consumption comes with JJA, TWS probably needs time to show all this influence in SON. The southwest rivers region (Yarlung Zangbo River, Nu River, and Lancang River) also showed significant TWS depletion, particularly in the upstream and downstream portions of the Yarlung Zangbo River. The area impacted by significant depletion was the largest in spring, while the depletion also became the most severe in autumn. Climate observation across this region proved that annual precipitation was decreasing over 11 years, with significant droughts in 2006, 2009, and 2012. Moreover, previous research also found ice loss in Himalaya from 2003 to 2007 (Matsuo and Heki, 2010).

Along with increasing precipitation in southeast China, there were significant increasing trends in TWS over the middle and lower reaches of the Changjiang River and southeastern coastal area. The main contribution to this significant increase occurred during the summer, when precipitation was the most concentrated of the year. There are two places that showed significant TWS increases in the Tibetan Plateau. One of these areas is around the Hoh Xil Mountains and the other is the headstream region of the Yellow River; they have maximum increasing trends of 2.59 and 1.77 cmyr-1, respectively. The Hoh Xil Mountains are located at the intersection of the inland lakes in the Qangtang Plateau and the north source of the Changjiang River. The headstream region of the Yellow River lies at the plateau's two largest freshwater lakes, Eling Lake and Zaling Lake. Previous research applied multi-source satellite data to reconstruct volume changes in the Tibetan Plateau's major lakes, and found that they showed similar spatial distribution to mass variations in GRACE during 2003–2010 (Song et al., 2013). According to satellite images, lakes in Hoh Xil overall showed a trend of expansion during 2000–2011. Further analysis suggested that increasing precipitation and decreasing evaporation were major factors contributing to this trend, and additional water recharge from melting glacier and frozen soil caused by climate warming was a minor factor (Yao et al., 2014; Duan et al., 2007). In the headstream region of the Yellow River, precipitation is the main recharge source to runoff, with a ratio of 63 %. Local observations revealed that there was an increasing trend in runoff as this region was becoming warmer and wetter during the period 2000–2012 (Lan et al., 2010, 2013; Wang et al., 2014). The Chinese government has launched an ecological protection and construction project in the Three-River Source region that started in 2005. According to monitoring data from the Qinghai Provincial Meteorological Bureau, average lake extents during 2005–2012 showed an increase of 34.7 and 64.4 km2, compared to those during 2003–2004 for Eling Lake and Zaling Lake, respectively.

In addition to the above large areas, there are also some other small regions showing strong TWS changes from 2003 to 2013. Along the northeast country border, there is significant TWS increase (0.34∼1.19 cmyr-1); this is mostly contributed by winter. The central part of the province of Jilin in northeast China shows severe TWS depletion in autumn (-1.10 ∼ -2.28 cmyr-1), but these linear trends have not passed the significance test (P value < 0.05). In northwest China, the unscaled GRACE data only show significant water depletion mainly around Tianshan Mountains, which is also identified with ice loss in previous research (Matsuo and Heki, 2010). However, trends from the scaled GRACE TWS also illustrate a significant TWS increase in the Turpan basin, while there is a depletion in its surrounding mountains. The Turban basin is the lowest basin in China, and Fig. 3b shows that the basin has a much smaller TWS amplitude than that in surrounding mountains. An extreme arid climate and local topography features in this region could make TWS more sensitive to climate change. Complex terrain in this region leads to a more complicated GRACE TWS signal mixture at a large spatial scale. Even though scale factors might have separated mixed TWS signals, limitations in factors' production should also be taken care of (Launderer and Swenson, 2012).