In this study, we use the simulations of monthly streamflow of China produced by six GHMs, namely DBH (Tang et al., 2007, 2008; Liu et al., 2016), H08 (Hanasaki et al., 2008a, b), LPJmL (Bondeau et al., 2007; Rost et al., 2008; Biemans et al., 2011; Schaphoff et al., 2013), MATSIRO (Takata et al., 2003; Pokhrel et al., 2015), PCR-GLOBWB (Wada et al., 2014), and WaterGAP2 (Flörke et al., 2013; Müller Schmied et al., 2014, 2016). Two experiments, i.e., simulations with (VARSOC) and without (NOSOC) human impact, were performed at a half-degree spatial resolution for the 1971–2010 period by using the six GHMs following the ISIMIP2a simulation protocol (https://www.isimip.org/protocol/\#isimip2a, last access: 18 February 2019). All the model runs used the same river routing map (DDM30; Döll and Lehner, 2002). For both experiments, the GHMs were forced by three global meteorological forcing products (GMFs), i.e., the PGMFD v.2 (Princeton; Sheffield et al., 2006), GSWP3 (http://hydro.iis.u-tokyo.ac.jp/GSWP3/, last access: 18 February 2019), and a combination of WFD (until 1978; Weedon et al., 2011) and WFDEI (from 1979 onwards; Weedon et al., 2014) datasets. Ensembles of annual streamflow are derived from the simulations of NOSOC (referred to as Qn) and VARSOC (referred to as Qv) experiments, respectively, for river segments (here a grid cell is treated as a river segment regardless of the cases where a grid cell contains several small river segments) which are then spatially averaged for individual basins. Long-term mean annual streamflow (MAF) in each river segment is calculated for both NOSOC and VARSOC simulations over a specific period (see Sect. 2.3) and is then spatially averaged over individual basins for each ensemble member. In addition to streamflow, total runoff from NOSOC and VARSOC simulations and water withdrawals from VARSOC simulations are also derived at grid cells and individual basins for associated analyses. The simulations may have large uncertainties over the Tibetan Plateau because long-term meteorological and streamflow observations are sparse in this region (Zhang et al., 2017) and the modeling of glacier melting is absent in most of the models. Therefore, the simulation data in the Tibetan Plateau region are removed and are not included in spatial averages by masking the grid cells with altitudes higher than 4000 m in all analyses.

Human impact considered in the VARSOC experiment (see the maps in Fig. S1 and Table S1 in the Supplement for more details) includes the time-varying areas for both irrigated and rain-fed cropland (Fader et al., 2010; Portmann et al., 2010) and reservoirs (dams) from the Global Reservoir and Dam (GRanD) database (Lehner et al., 2011) including their commissioning year (see Fig. S1 and Table S1). Reservoir regulation was considered in the VARSOC experiment, which often reduces high streamflow in high-flow seasons and increases streamflow in dry seasons (Masaki, et al., 2017). Interbasin water transfer was not considered in any of the model runs. The simulations of water withdrawals are different between the GHMs with respect to water use requirements and water withdrawal sources which are shown in Table S1. The sources of water withdrawals, depending on models, may include river channel, reservoirs, groundwater, and lakes, and their fractions may be determined from reported statistics (e.g., Siebert et al., 2010) or estimated in models (Wada et al., 2014). In addition to the irrigation water requirement which is usually estimated by coupling crop models, most GHMs considered the requirements for domestic and industrial water use that were prescribed in H08 (Hanasaki et al., 2008a), LPJmL, and MATSIRO (Pokhrel et al., 2015) or were estimated according to the population; socioeconomic and technological development in PCR-GLOBWB (Wada et al., 2014); and the population, thermal electricity production, gross added value, and technological change in WaterGAP (Flörke et al., 2013). The water use requirement for livestock was also prescribed in the LPJmL model and estimated according to livestock densities in PCR-GLOBWB and WaterGAP2.

The ISIMIP2a streamflow simulations have been extensively validated with observations over the world in several studies (Liu et al., 2017b; Veldkamp et al., 2018; Zaherpour et al., 2018) but were not fully evaluated in China due to limited observations, particularly for the water withdrawals. Therefore, before the quantitative attribution, an evaluation of the multimodel simulations is performed, which may add confidence regarding the GHM performance over China. Observations of monthly streamflow from 44 hydrological stations in China (Fig. 1) during 1971–2000 are used for model validation. The observations since 2001 are not available in this study. Some stations are relocated on the map to reconcile the catchment areas of the stations and the accumulative flow areas of corresponding gird cells from the DDM30 river network. After relocation, the differences are mostly less than 10 % (about 50 % at five stations) between the reported catchment areas of stations and the accumulative flow areas from the DDM30 river network. Annual water withdrawals in individual basins for the years 1980, 1985, 1990, 1995, and 1997–2010 were collected from China Water Resources Bulletin from the Ministry of Water Resources (MWR) of China (http://www.mwr.gov.cn/sj/tjgb/szygb/, last access: 18 February 2019).

The study period is evenly split into two sub-periods (P1 for 1971–1990 and P2 for 1991–2010). The DHI-induced MAF changes over time is calculated as QhP1=100×QvP1-QnP1QnP1QhP2=100×QvP2-QnP2QnP2, where QhP1 and QhP2 denote MAF changes (%) induced by DHI during the sub-periods P1 and P2, respectively; QvP1 and QvP2 denote MAF from the VARSOC experiment for the two sub-periods, respectively; and QnP1 and QnP2 denote MAF from the NOSOC experiment for the two sub-periods, respectively.

The contribution of DHI change (corresponding to climate variability) on streamflow changes between the two sub-periods is also examined. MAF difference between the two periods in the VARSOC experiment is defined as the total MAF changes (ΔQa) caused by both climate variability and DHI change from P1 to P2, which is expressed as a percentage of the MAF of the first sub-period P1: ΔQa=100×QvP2-QvP1QvP1. The difference between the two periods in the NOSOC experiment is defined as streamflow changes induced by only climate variability (ΔQc) and expressed as a percentage of QvP1 in order to be comparable with ΔQa: ΔQc=100×QnP2-QnP1QvP1. The difference between ΔQa and ΔQc then counts as MAF changes induced by DHI change (ΔQh) between the two sub-periods: ΔQh=ΔQa-ΔQc=100×QvP2-QvP1-QnP2-QnP1QvP1. Unless otherwise stated, ΔQa, ΔQc, and ΔQh are relative changes (%) with respect to QvP1 in this paper.

To address the potential uncertainty resulting from the use of sub-periods, similar analyses are performed for three different sub-periods, namely 1981–1990, 1991–2000, and 2001–2010, with comparison to the sub-period 1971–1980. For these analyses, MAF is calculated over each decade.

In addition to streamflow, changes in water withdrawals and total runoff between the two sup-periods are also analyzed to explore their links with MAF changes.

Ensemble medians across the 18 GHM-GMF combinations (six GHMs and three GMFs) are used for analyses of streamflow and runoff. But 12 ensemble members are used for water withdrawals because only four GHMs (H08, LPJmL, PCR-GLOBWB, and MATSIRO) provide related output for the ISIMIP2a simulations. The interquartile range (IQR), i.e., the range between 25th and 75th percentiles, is calculated to present the spread across multimodel ensembles. The ratio of IQR to the median is used to measure the uncertainty in multimodel simulations of streamflow, which is comparable across regions.

In this study, the northern basins refer to the Songhua River (SH), Liao River (LR), northwestern rivers (NW), Hai River (HA), Yellow River (YR), and Huai River (HU); the southern basins refer to the Yangtze River (YZ), southeastern rivers (SE), southwestern rivers (SW), and Pearl River (PR; Fig. 1). The ensemble medians of MAF at grid cells over the 1971–2000 period from the VARSOC experiment show distinct spatial pattern of high streamflow in southern basins and relatively low streamflow in northern basins (Fig. 1). The multimodel simulations show larger spreads in the northern basins. The ratios of IQR / median are larger than 1 or 2 in the northwestern basin, the Yellow River basin and Liao River basin. Smaller spread (IQR / median less than 0.5) is found in the middle and lower reaches of the Yangtze River basin, the Pearl River basin, and the southeastern basin.

The inner plot shows the comparison between observed seasonal streamflow averaged across all hydrological stations and the averaged simulations in all river segments identified by stations over the 1971–2000 period. The ensemble medians of seasonal cycle generally coincide with the observations. However, there are large variations across all model ensembles, with some of them deviating from observations. It should be noted that the stations are located at different reaches of individual basins. Thus, the station-averaged estimates are largely dominated by those with large streamflow (e.g., at the lower reaches). Additionally, the coverage of stations used is relatively small (due to data availability), especially in hydrologically variable regions like in the northwestern rivers, leading to an evaluation of the performance of the GHMs in the whole basin that is not necessarily representative. The model spreads in the ensembles of seasonal streamflow in the northern basins are relatively larger than those in the southern basins (see Fig. S2 for each basin). Comparison between the simulated and observed annual streamflow (Fig. S3) shows similar patterns to the seasonal streamflow with respect the discrepancies between northern and southern basins. The Nash–Sutcliffe coefficient was calculated for the multimodel median and observed monthly streamflow at each station (see Table S2), which shows that the multimodel medians have better performance in the southern basins. This evaluation indicates that the multimodel simulations have relatively poor performance in northern basins and most stations with low Nash–Sutcliffe coefficients have smaller streamflow (e.g., in dry areas or upper reaches). The large spreads between models underline the necessity of using ensemble medians rather than individual models for the attribution of streamflow changes.

Compared to the reported data by the MWR of China, the ensemble medians from ISIMIP2a simulations underestimated water withdrawals in most northern basins except for the Yellow River (Fig. 2). The simulations underestimate water withdrawals by more than 50 % in the northwestern rivers and the Hai River and by more than 30 % in the Songhua River and Liao River. The simulated water withdrawals are 12 % less than reported data in the Huai River. In the Yellow River and the southeastern rivers, water withdrawals are overestimated by 20 % or more. The overestimation of water withdrawals is the largest (80 %) in the southwestern rivers. Small differences between simulations and reported data are found in the Yangtze River (-1 %) and the Pearl River (-6 %). The large deviations in the multimodel simulations of water withdrawals could make the modeling of streamflow more challenging (Döll et al., 2016; Wada et al., 2017).

Figure 3 shows the spatially averaged ensemble medians of Qn and Qv over China, the northern basins and the southern basins, respectively. Qn and Qv show considerable annual variations and no statistically significant trends over the 1971–2010 period. The relative differences between Qn and Qv over China range from -8 % to -4 % and show a statistically significant downward trend over the study period (Fig. 3a). The differences between Qn and Qv over the northern basins are larger than those for the southern basins. The absolute differences (not shown here) are -37 to -14 (m3 s-1) for the northern basins and are -37 to -7 (m3 s-1) for the southern basins. The relative differences for the northern basins (Fig. 3b) are also larger than those for the southern basins (Fig. 3c). The former ranges from -30 % to -10 %, while the latter ranges from -4 % to -1 %. A statistically significant downward trend is found in the relative differences for the northern basins, while a non-significant downward trend is found for the southern basins. The downward trend in the differences indicates that annual streamflow has been increasingly affected by human impact.

Considerable decreases in long-term MAF are induced by DHI in the two sub-periods (Fig. 4a and b for QhP1 and QhP2, respectively) in most northern basins. About 3 % and 4 % of total river segments in China show large negative values (i.e., less than -30 %) of QhP1 and QhP2, respectively, which are mostly found in some parts of the northwestern rivers and the North China Plain. QhP1 and QhP2 are negative for more than 90 % of the river segments and range from -5 % to 0 in more than 60 % of the river segments of China. The magnitudes of the basin-averaged QhP1 and QhP2 are larger than 10 % in northern basins except for the Songhua River (Fig. 4c). The magnitudes of QhP2 are larger than QhP1 for all basins, especially in the Yellow River. The northwestern rivers show the largest negative values of QhP1 and QhP2 (-31.6 % and -33.5 %, respectively), which is followed by the Hai River (-24 % and -25 %), the Yellow River (-17 % and -21 %), and the Huai River (-17 % and -19 %). DHI induced slight decreases in MAF (-3.4 % to -0.3 %) in the southern basins. Overall, DHI altered MAF by -4.4 % and -5.6 % in China during the sub-periods P1 and P2, respectively.

The MAF changes induced by DHI change and climate variability between the two sub-periods are shown in Fig. 5. In general, total MAF changes (ΔQa; Fig. 5a) are larger in northern basins, except the Songhua River, than in southern basins. Compared to the first sub-period, in the second sub-period MAF increased by more than 30 % in many river segments of the northwestern rivers and increased by more than 5 % in large parts of the Huai River, which may be due to the return flow from water withdrawals. MAF increases are also found in considerable areas of southern basins such as the Yangtze River and the southwestern rivers. MAF decreases are found in most river segments in the Yellow River, the Hai River, and the Liao River. Significant negative values of ΔQa (less than -20 %) are found in some river segments in the upper reaches of the southwestern rivers and some parts of the northwestern rivers. The total MAF decreased by more than 10 % (ΔQa<-10 %) and increased by more than 10 % (ΔQa>10 %) in about 24 % and 17 % of river segments of China, respectively.

MAF changes induced by climate variability between the two sub-periods (ΔQc; Fig. 5b) have very similar spatial patterns to ΔQa (Fig. 5a). This indicates that climate impact dominates MAF changes during the two sub-periods. The magnitudes of ΔQc are relatively smaller than those of ΔQa in the Hai River and the Yellow River but are larger in the northwestern parts of the northwestern rivers. MAF changes induced by DHI change (ΔQh; Fig. 5c) are generally large and negative in northern basins. A decrease larger than 10 % in the MAF induced by DHI is found in some segments of the northwestern rivers and the lower reaches of the Huai River, the Hai River, and the Liao River. Positive values of ΔQh are small and are mostly found in southern river segments. Climate impact dominated MAF changes in most river segments (88 %) of China (Fig. 5d). Only 12 % of river segments show MAF changes that are mainly caused by DHI change, which are mostly in the northern basins.

MAF changes induced by climate variability (ΔQc) versus those induced by DHI change (ΔQh) for all river segments are shown in Fig. 6a. Note that very few river segments with values of ΔQc (0.9 % of total river segments) and ΔQh (0.4 %) beyond [-100,100] are not shown in the figure. Magnitudes of ΔQc are much larger than those of ΔQh. The latter ranges -5 % to 5 % in most (∼ 81 %) river segments. ΔQh is less than -10 % in only about 7 % of river segments of China, while even fewer (∼ 3 %) segments show ΔQh values larger than 5 %. The values of ΔQc range from -10 % to 10 % in more than half of river segments and range from -20 % to 20 % in nearly 80 % of river segments (see Table S3 for related numbers). ΔQh is negative in 70 % of river segments, while negative values of ΔQc are found in more than half of the river segments of China (see the percentage numbers in Fig. 6a and Table S3).

The total MAF spatially averaged over China decreased by only 1 % from the first sub-period to second sub-period (Fig. 6b; see also Tables S4–S6 for more details of spatially aggregated ensemble members and medians of ΔQa, ΔQc, and ΔQh in basins). At the basin scale, the magnitudes of MAF changes are usually very small (less than 2 %) in southern basins and are relatively large in northern basins (5 % to 13 %). ΔQa in the Hai River shows the largest decrease of 13 %, which is followed by a nearly 10 % decrease in the Yellow River and a 7 % decrease in the Liao River. Increases in total MAF are found in the northwestern rivers (10 %), the Huai River (1.8 %), the Pearl River (1.3 %), and the southwestern rivers (1.2 %), which are consistent with the spatial patterns shown in Fig. 5a. DHI change causes decreases in MAF (negative ΔQh) in all the basins, resulting in a larger decrease or a smaller increase in ΔQa compared to ΔQc. The largest negative values of ΔQh are found in the northwestern rivers (-8 %), the Huai River (-5.4 %), and the Hai River (-4.6 %; see Tables S4–S6). ΔQh is about -2.6 % for the Liao River and the Yellow River. ΔQh is only about -0.7 % to -0.07 % in southern basins. The increase of MAF induced by climate variability (ΔQc) is the largest in the northwestern rivers (18 %), followed by the Huai River (6 %) and the Pearl River (1 %), and climate variability caused a nearly 9 % decrease in MAF in the Hai River and the Yellow River.

For both sub-periods, the estimates of long-term mean annual water withdrawals are large (more than 100 × 106 m3 yr-1) in many areas of the Huai River, the Hai River, and the Yellow River (Fig. 7a). Large water withdrawals are also found in some lower reaches of the Yangtze River. In these regions, mean annual water withdrawals are usually larger in the lower reaches compared to the upper reaches and significantly increased from 1971–1990 to 1991–2010. The relative changes in water withdrawals between the two sub-periods show distinct spatial patterns from northern to southern basins and generally increased at all river segments of China (Fig. 7c). The spatial patterns of changes in water withdrawals resemble those of ΔQh, with large values in the Huai River, the Hai River, and the Yellow River, but they are relatively smaller in the northwestern rivers. Similar analysis is performed for changes in total runoff to examine its linkage with streamflow changes. The spatial patterns of changes in total runoff induced by DHI change between the two sub-periods (Fig. 7d) are different from that of ΔQh (Fig. 5c). Total runoff changes are positive in most areas of China due to increasing irrigation water (from both local and external sources) which partly becomes return flow, especially in the northwestern rivers. Large changes are also found in upper and middle reaches in the Yellow River, the Liao River, and the Hai River. The magnitude of changes is less than the that induced by climate variability (not shown here), which is similar as Fig. 5d. This indicates that the runoff changes are less linked to streamflow changes in the study period.

DHI causes MAF decreases in both of the sub-periods. At the basin level, DHI resulted in decreases by one-fifth to one-third of the long-term MAF based on Qn in northern basins and slightly altered MAF in southern basins of China. The spatial patterns of the MAF altered by DHI (Qh) are generally in accordance with those reported by previous studies (Liu and Du, 2017) and are very close to those of irrigated areas of China (see Fig. S1). The expansion of agriculture and enhanced irrigation and food demands should be the main reason for the large DHI on streamflow in northern basins (Liu et al., 2015; see also Fig. S1c), where agricultural water use accounts for about 70 %–90 % of total water use as reported by China Water Resources Bulletin from 1997 to 2010. Water withdrawal for irrigation is less due to the large streamflow and relatively wetter conditions in southern basins. Limited water resources can further amplify the effects of damming on river segments in northern basins (Yang and Lu, 2014), despite having fewer reservoirs compared to southern basins (see Fig. S1a).

Though MAF changes between the two sub-periods are relatively small, especially in southern basins, the respective contributions of climate variability and DHI change are still distinguishable. In general, streamflow changes are dominated by climate variability between the two sub-periods in most river segments of China. The small portion (12 %) of river segments where DHI change outweighs climate impact on MAF changes are mostly in northern China. The small magnitudes of MAF changes induced by DHI change between the two sub-periods may be partly due to the fact that DHI change is not significant in most areas of China in the VARSOC experiment. Although the irrigated areas in both the northern and southern basins increased by about 20 % in the second sub-period (see Fig. S1c), the changes between the two sub-periods are small (less than 5 %) in many areas except in the Huai River and the Hai River (see Fig. S1b). Furthermore, there are only a few reservoir data from the GRanD database after the year 2000, and most reservoirs in China were built in the first sub-period (see Fig. S1d). The reservoirs lacking construction years were set to be built (and operated) at the beginning of the experiment in the model runs.

It is noted that the absolute MAF changes between the two sub-periods are large in main streams in both southern and northern basins (see Fig. S4); the significant MAF changes induced by DHI change in the Yangtze River are associated with the large reservoir regulations, e.g., the Three Gorges reservoir (Wang et al., 2013a).

Overall, the spatial patterns of water withdrawal changes (Fig. 7c) are similar to MAF changes induced by DHI change (ΔQh; Fig. 5c) between the two sub-periods. Though water use partly infiltrates into land surface and eventually increases local runoff (see Fig. 7d), water withdrawals should be the major DHI that contributes to decreases in streamflow in most river segments in China. For example, the significant decreases in MAF are supposed to be largely related to water withdrawals in the northwestern rivers where streamflow is low and only one reservoir was included in the VARSOC simulations. The water withdrawal changes in northwestern rivers are relatively small compared to other northern basins, but they still have significant implications because of the limited water resources. As mentioned above, water withdrawal for agricultural irrigation accounts for the largest proportion of human water use in China, most of which finally evaporates into the atmosphere through both crop and soil because of the low irrigation efficiencies (Zhu et al., 2013), which might be the main source depleting the streamflow and local water resources. Though the return flow might increase runoff over most river segments of China (Fig. 7d), it seems to be only a small proportion of the water withdrawals and does not offset the decreases in streamflow. Unlike water withdrawals, the effects of reservoir regulation on annual streamflow are mixed in current GHMs, as reservoir regulation generally reduces streamflow in flood (and growing) seasons while streamflow increases in dry seasons (Masaki et al., 2017).

Though the effects of DHI change on streamflow are smaller compared to those of climate variability in China (see Sect. 4.2), the DHI-induced streamflow changes significantly increased, particularly in the northern basins over the 1971–2010 period (Fig. 3). The northern basins have relatively poor water endowments and have been identified as regions that are highly sensitive to climate change (Piao et al., 2010). The relatively high DHI further increases the pressure on and threats to water management and adaptation to future climate change in these regions. For example, frequent zero flow was observed in some reaches of the Yellow River due to climate variability and human water use in the 1990s (Tang et al., 2013). Most northern regions suffered severe water scarcity during the past decades (Liu et al., 2017a), and the water resources have been increasingly insufficient for human water needs in many areas of northern basins (Liu and Xia, 2004). The unregulated pumping of non-renewable groundwater has resulted in significant depletion and far-reaching effects on both hydrological cycle and human society in these regions (Feng et al., 2013). The DHI change over time further enlarges associated streamflow changes in these basins (see Fig. 4c and S5). The situation could be worse if no adaptation is taken to act under future climate change (Piao et al., 2010; Liu et al., 2015). Thus, in view of the considerable DHI in these regions, there is an urgent need for a structural transformation of the economy towards reducing water use and a sustainable development.

The major uncertainty in this quantitative assessment usually originates from input forcings (Müller Schmied et al., 2014) and inter-model differences such as human impact parameterizations (Liu et al., 2017b). That is, the uncertainties in streamflow simulations would spread to the assessment. For example, there are very few meteorological observations in the northwestern rivers, possibly leading to considerable uncertainties in the meteorological forcings used to drive GHMs. Furthermore, the GHMs cannot fully reflect sectoral water withdrawals (Fig. 2; see also Huang et al., 2018) because of a lack of data on water abstractions for human use from surface and groundwater sources (Liu et al., 2017b). Meanwhile, the different water withdrawal requirement and withdrawal sources considered in GHMs (see Table 1) may result in inter-model uncertainty in the estimates of water withdrawals and perhaps enlarge the discrepancy in streamflow simulations. The multimodel ensemble medians seem to be in line with observations averaged across the stations in China, but large discrepancies are found in some basins (Fig. S2). This indicates a large potential for the GHMs to improve streamflow simulations in China. It should be noted that we have relocated some stations on the map to reconcile the catchment areas of the stations and the corresponding grid cells on the DDM30 river network. However, catchment areas still are inconsistent between some stations and their corresponding grid cells, especially for the stations not on the main stream. This may be partly responsible for the deviation between simulated and observed streamflow. More hydrological observations (from large catchment areas) are necessary to perform a comprehensive evaluation of streamflow simulations.

In addition to the uncertainties in multimodel simulations of streamflow, the quantitative assessment depends on the selection of comparison periods. To examine the possible effects of the selection of sub-periods, we perform similar assessments for different sub-periods, i.e., MAF changes in three decades, 1981–1990, 1991–2000, and 2001–2010, compared to the first decade (1971–1980). The assessments show similar patterns of MAF changes as in Fig. 5, with larger relative changes in most northern basins (see Fig. S5 for the analysis at basin scale). Effects of climate variability on streamflow vary over different sub-periods. In contrast, DHI change usually resulted in MAF decrease across all basins, and this impact slightly increases over time (see Table S7 for corresponding numbers), especially in the northern basins such as the Yellow River, the northwestern rivers, the Liao River, and the Hai River. In the Yellow River, MAF changes induced by DHI change outweigh those induced by climate variability in the 2001–2010 period. Human activities may be weaker in China before the year 1971, and the DHI change could be larger if compared to earlier periods (e.g., Müller Schmied et al., 2016). This assessment suggests that the magnitudes of the impacts of both climate variability and DHI change on streamflow are associated with specific sub-periods; however, DHI change decreased streamflow in almost all basins in the study period.

Both this study and previous ones (Table S8) show that DHI (change) almost always contributes to decreases in streamflow in China, but the DHI contributions are much more significant in previous assessments compared to this one. Previous studies have shown that DHI contributed to decrease in streamflow by 20 % to 80 % across catchments in the Hai River, Yellow River, and Huai River (see Table S8; it should be noted that the proportions in the table were calculated as 100×ΔQh/ΔQa). In four cases the DHI contributions are larger than those of climate impact, and in most cases DHI contributes more than 40 % in these studies (see Table S8 for the results from previous studies), while DHI contributions are mostly smaller than climate variability in this assessment (Fig. 6a). There are several reasons for the large differences between this assessment and previous ones which make their results not comparable directly, such as different methods and data, sub-periods, and study areas (see Table S8 for details). Unlike this study, the previous assessments were usually performed in small catchments that experienced evident human activities and comparison periods were usually chosen using statistical approaches (e.g., abruptly changing point detection for a time series).

One major difference between previous studies (e.g., Li et al., 2007; Bao et al., 2012) and this study is that the former estimates DHI contribution by comparing simulations with observations, while we compare two simulation experiments. The former may be subject to uncertainty in comparing the data from two systems (i.e., the model and the real world). In this study, the two simulation experiments favor the estimation of DHI contribution in a consistent manner that is largely free of uncertainty in the data from different systems. The multimodel simulations also allow profiling the uncertainties among models and input forcings, which is difficult for a single model assessment. However, the deficiency of this approach is that DHI is not real. Therefore, the assessment is inevitably influenced by the extent to which the models can reproduce the real DHI. Considering the complexity of DHI on streamflow and the ability of current hydrological models in reproducing historical hydrological changes, multimodel simulations and different attribution approaches are well worth obtaining more robust assessments (Liu et al., 2017b; Yuan et al., 2018).