In the YRB, the area above the Tangnaihai (TNH) hydrological station (100∘09′ E, 35∘30′ N) is defined as the headwater region. The Toudaoguai (TDG) station (111∘04′ E, 40∘27′ N) is the demarcation point between upper and middle reaches. The region between the TDG and Huayuankou (HYK) gauges (113∘39′ E, 34∘55′ N) is the middle reach where the region between the TDG and Longmen (LM) gauges (110∘35′ E, 35∘40′ N) is the main sedimentation formation area of the YRB. The study area is the catchment above the HYK station with a drainage area of 730 036 km2 (∼97 % of the total area of the YRB), and the mean annual runoff in the study area accounts for ∼98 % of that in the whole YRB (Tang et al., 2013). Areas of contribution for TNH, TDG, LM, and HYK are approximately 121 972, 367 898, 497 552, and 730 036 km2, respectively. The study area is divided into four subregions (source region, TNH–TDG, TDG–LM, LM–HYK) between the target gauge and the adjacent upstream gauge from the TNH gauge to the HYK gauge, as illustrated in Fig. 1.

The observed daily data from 265 meteorological stations, including the daily time series of precipitation, maximum temperature, minimum temperature, and wind speed from 1980 to 2018, were obtained from the China Meteorological Administration (http://data.cma.cn/, last access: 10 October 2019). We calculated the daily mean temperature by averaging daily maximum and minimum temperatures. The 8 d time series of vegetation leaf area index (LAI) at 500 m from 1982 to 2018 used in this study was obtained from The Global Land Surface Satellite (GLASS) product (Xiao et al., 2014) (http://glass-product.bnu.edu.cn/, last access: 7 January 2020). We obtained land cover data for every 5 years during 1985–2020 from the GLC_FCS30 product (Zhang et al., 2021), which was the first global land cover product with a fine classification system at 30 m (http://www.geodata.cn/, last access: 4 November 2021). Elevation data obtained from the Shuttle Radar Topography Mission (SRTM) digital elevation dataset at 90 m (https://www.gscloud.cn/, last access: 7 August 2019) were used to delineate river networks that are necessary for runoff routing of the hydrological model. The soil texture data were derived from the 1 km China soil map based on the harmonized world soil database (HWSD) (v1.1) (http://data.tpdc.ac.cn/en/, last access: 15 September 2019). The China terrace proportion map at 1 km resolution (Cao et al., 2021) in 2018 was downloaded from https://doi.org/10.5281/zenodo.3895585. The global surface water product at 30 m from 1984 to 2020 was available from the Joint Research Centre (JRC) (https://global-surface-water.appspot.com/download, last access: 20 May 2021).

For runoff, there are four mainstream gauges shown in Fig. 1. Monthly naturalized runoff from 1980 to 2018 was provided by the Yellow River Conservancy Commission of the Ministry of Water Resources. Naturalized runoff at the target gauge was estimated by adding data of human water use from irrigation, industrial, and domestic sectors over the drainage area of the target gauge back to the observed runoff at the target gauge (Yuan et al., 2017). We used naturalized runoff to calibrate the hydrological model for simulating natural hydrological processes.

We used the slope of the simple linear regression (Wang et al., 2022) to characterize the interannual change trend of streamflow and influencing factors including precipitation, temperature, wind speed, and LAI over the YRB. The t test was used to examine the significance level of this trend. In addition, the contribution of monthly streamflow change at a given month to the annual streamflow change was also determined by dividing the trend of the monthly streamflow by the trend of the annual streamflow.

Since changes of the intra-annual temporal pattern in the precipitation and LAI are also able to affect the annual streamflow, we have taken the ratio of observed monthly to annual precipitation or LAI as the indicator of the intra-annual temporal pattern in this study, and its change trend of each month was also analyzed. To explore more details on the relationship between the temporal variability of precipitation and streamflow, double mass curve (Zhang et al., 2011) was performed to detect the abrupt change point and baseline period in the annual streamflow time series (Mu et al., 2007; Gao et al., 2011).

The VIC model uses the variable infiltration curve (Liang et al., 1994) to account for the spatial heterogeneity of runoff generation. It assumes that surface runoff for the upper two soil layers is generated by those areas where precipitation exceeds the storage capacity of the soil. The methods from the ARNO model (Todini, 1996) were used to describe base flow generation which only happened in the third soil layer. A separate routing model was then coupled with the VIC model to simulate streamflow (Lohmann et al., 1998), where the runoff generated in each grid cell is routed to selected points through the channel network.

To balance the high cost of computation and the characterization of heterogeneous underlying surface, we performed simulations using the VIC model on a 0.1∘×0.1∘ grid scale at a daily time step. The inputs of the VIC model include meteorological forcings, vegetation parameters, land cover, and soil parameters. The meteorological forcings were derived by interpolating gauged daily precipitation, maximum and minimum temperatures, and wind speed from stations into a resolution of 90 m based on the AUSPLINE software and DEM data; we then calculated the spatial average of interpolated data within a grid cell, as illustrated in Fig. 2. The default setting of the VIC model only considers the climatology of vegetation (e.g., 12-month LAI), and the monthly LAI and land cover are stationary in each year during the simulation period. Therefore, the impacts of continuous interannual change of LAI and land cover types on hydrological processes have rarely been discussed in previous studies using the VIC model (Xie et al., 2015; Yang et al., 2019; Zhai and Tao, 2021). In this study, the simulation scheme of the VIC model (version 4.1.2.a) considering time-variant LAI was designed as the following two steps:

Step I: The GLC_FCS30 product was firstly resampled to the same resolution (500 m) of the LAI product. Owing to a lack of yearly land cover data, the land cover data from the GLC_FCS30 product in the ith year were used to represent the land cover from i(-4)th year to the ith year. We smoothed the 8 d LAI time series with the adaptive Savitzky–Golay filter (Chen et al., 2004) to eliminate the abnormal LAI contaminated by cloud, signal errors from sensor, etc. The smoothed 8 d LAI was then aggregated to a monthly value with temporal averaging for each year. Finally, the area fractions and average monthly LAI value for each land cover type in each 0.1∘×0.1∘ grid cell in each year were calculated respectively (Fig. 2).

The objective of this study was to investigate the contributions of changes in climate and vegetation to runoff changes, rather than to simulate runoff accurately from 1982 to 2018. Therefore, we adopted the baseline period to calibrate the seven numerical soil parameters, including the infiltration parameter b, the depths of three soil layers (d1, d2, and d3), and the three parameters in the base flow scheme (Dm, Ds, and Ws) (Xie et al., 2007; Shi et al., 2008), in different subregions.

To find the optimal parameter set, an optimization algorithm of the multi-objective complex evolution of the University of Arizona (MOCOM-UA) from Yapo et al. (1998) was implemented, and the Nash–Sutcliffe efficiency (NSE), relative bias (Bias) and root mean square error (RMSE) were used as the objective function to assess the model performance as illustrated in Eqs. (1)–(3). The automatic calibration was carried out by running the VIC model thousands of times during the calibration period (1980–1993), of which the first 2 years (1980–1981) were used for warm up, and the period 1994–1999 was the validation period. 1NSE=1-∑i=1NQobs,i-Qsim,i2∑i=1NQobs,i-Qobs‾, 2Bias=∑i=1NQsim,i-∑i=1NQobs,i∑i=1NQobs,i,3RMSE=1N∑i=1NQsim,i-Qobs,i2, where Qsim and Qobs are the simulated and observed monthly streamflow, respectively, Qobs‾ is the arithmetic mean of the observed monthly runoff, i is the ith month, and N is the total number of months in the calibration period.

This study assumes that the same amount of relative bias of the annual streamflow trend during the calibration period will be transformed to the scenario simulation, and this relative bias was then deducted when calculating impacts of climate and vegetation on runoff (Luan et al., 2020). In this way, we can minimize the impact of hydrological simulation error in attributing the annual streamflow change trend.

The naturalized annual streamflows of the HYK station, source region, TNH–TDG, TDG–LM, and LM–HYK are provided in Fig. 3a. A significant decreasing trend was observed from the annual streamflow time series of the HYK station during 1982–2018, with a negative trend of -3.71×108 m3 yr-1. Spatially, all subregions reported downward annual streamflow trends, with different contributions of 20.7 %, 20.6 %, 14.6 %, and 44 % on the annual streamflow trend of HYK from the source region to LM–HYK. Temporally, all monthly streamflow experienced negative trends at the HYK station, with the greatest reduction (18.6 %) being found in August. Most monthly streamflow trends of the four subregions were negative, and the greatest contributions of monthly to annual trend occurred during July for the source region, August for TNH–TDG and TDG–LM, and in September for LM–HYK.

Runoff coefficients characterizing runoff-yield capacity were calculated as streamflow divided by precipitation for all the subregions as illustrated in Fig. 7a. It was found that overall trends in all the subregions were negative during 1982–2018, although there were short-period upward trends from 2000 to 2018 in all subregions excluding TDG–LM. To detect the abrupt change time of the relationship between precipitation and streamflow, the cumulative curves of precipitation and streamflow of the four subregions were calculated and plotted (Fig. 7b). The discrepancy of cumulative precipitation and streamflow observed from Fig. 7b indicated that the stationary precipitation–streamflow relationships have changed in all subregions, and the deviation of the streamflow from precipitation is more significant in the middle reaches than in the upper reaches. It has been noted that significant abrupt changes in different subregions occurred in the same year of 1999. Thus, the study period was divided into two periods: 1982–1999 and 2000–2018. It is evident from Fig. 7c that the precipitation–runoff relationship changed significantly between these two periods, and the regression line of precipitation and runoff during 1982–1999 was always above that during 2000–2018, suggesting that runoff in the period 2000–2018 was significantly reduced when the same precipitation in the period 1982–1999 occurred. Therefore, it is reasonable to split the entire period into these two short periods. It could be concluded that the relationship between the annual precipitation and streamflow presents a non-stationary state in the YRB from 1982 to 2018.

According to the calculated abrupt change point from the precipitation–streamflow double mass curves in Fig. 7, the period 1982–1999 was defined as the reference period, of which the calibration and validation periods for calibrating the VIC parameters were 1982–1993 and 1994–1999, respectively. The seven model parameters calibrated in different drainage areas are shown in Table 3. The monthly hydrographs and average seasonal cycles of the simulated and naturalized streamflows for different catchment regions are shown in Fig. 8, and the accuracy metrics of all simulations in Fig. 8 are summarized in Table 4. According to the VIC simulations at HYK, the monthly NSE, RMSE, and Bias are 0.89, 387.4 mm, and -1.6 % for the calibration period and 0.8, 386.6 mm, and 6.9 % for the validation period. Averaged across all four catchment regions, the monthly NSE is 0.69, RMSE is 171.9 mm, and Bias is 5 % during the calibration period, and the monthly NSE is 0.6, RMSE is 156.8 mm, and Bias is 9.5 % during the validation period (Table 4). As per the performance criteria given by Moriasi et al. (2007), simulation results indicate that the VIC model has a good performance in simulating hydrological processes for subbasins and subregions. In addition, Table 4 also shows the multiyear average monthly streamflow during 1982–1999, and the NSE is larger than 0.85 in all catchment regions, except for TDG–LM, thus indicating that the seasonal cycles of streamflow can also be perfectly captured by the VIC simulation.

The impacts and relative impact rates of eight influencing factors on the annual streamflow trends in different drainage areas were calculated using Eqs. (6)–(14), as illustrated in Fig. 9 and Table 5. From 1982 to 2018, the annual streamflow trend at HYK was -3.71×108 m3 yr-1, of which changes in interannual precipitation (P_inter), temperature (T_inter), wind speed (WS_inter), intra-annual temporal pattern of precipitation (P_intra), interannual LAI (LAI_inter), intra-annual temporal pattern of LAI (LAI_intra), interactive effects of climate variables and vegetation (Interactive), and residual underlying surface (Resi.) accounted for 15.1 % (1.14×108 m3 yr-1), -23.5 % (-1.77×108 m3 yr-1), 8.7 % (0.66×108 m3 yr-1), 1.4 % (0.1×108 m3 yr-1), -26.6 % (-1.99×108 m3 yr-1) and -6 % (-0.45×108 m3 yr-1), -3.5 % (-0.26×108 m3 yr-1), -15.2 % (-1.14×108 m3 yr-1), respectively. For the HYK station, the contributions of all climate variables to the streamflow trend were positive, except temperature, while larger negative effects from the underlying surface change offset the slight positive effects of climate change on the streamflow trend (Fig. 9). It is concluded that vegetation change was the dominant driving factor for the long-term decreasing trend of streamflow from 1982 to 2018 in the YRB. Meanwhile, the effects of non-vegetation underlying surface changes (e.g., water and soil conservation engineering measures, permafrost melting) on reducing streamflow cannot be ignored.

Due to the divergent change of climate variables and underlying surfaces, the impact of different influencing factors on the streamflow trend in different subregions exhibited obvious spatial variability. The net total effect from interannual changes of all climate variables exhibited a negative influence on the streamflow increase for all subregions, except for the TDG–LM subregion which showed a positive impact of 0.47×108 m3 yr-1. The contribution of temperature on the decreasing trend of streamflow in the upper reaches is greater than that in the middle reaches. Contributions of the intra-annual temporal pattern change of precipitation on the streamflow trend illustrated obvious spatial heterogeneities. The impact of this factor was positive in the source region (2.1 %) and LM–HYK (11.2 %), whereas its negative effects were observed for the TNH–TDG (-2.6 %) and TDG–LM (-9.9 %) subregions. It was found that not only the interannual increase of LAI, but also the intra-annual LAI temporal pattern change had effects of reducing streamflow. Direct total impacts from vegetation change were negative for streamflow trends and accounted for -10 % (-0.19×108 m3 yr-1), -19.5 % (-0.28×108 m3 yr-1), -36.6 % (-0.72×108 m3 yr-1), and -46.9 % (-1.25×108 m3 yr-1) of the streamflow trends in the source region, TNH–TDG, TDG–LM, and LM–HYK, respectively. Compared with direct effects of vegetation change, the two-way interactive effects of vegetation and climate variables were relatively low in all subregions. The impacts of residual underlying surface change were comparable to that of vegetation greening, with a maximum contribution of -21.4 % (-0.41×108 m3 yr-1) occurring in the source region.

Previous studies have suggested that precipitation is the main factor controlling runoff change with climate change (Dan et al., 2012; Wang et al., 2016; Liu and Liu, 2018). In this study, we further found that the simulated annual streamflow trend could be changed by not only interannual precipitation (S2–S1) but also intra-annual monthly to annual precipitation ratio (S3–S2), which indicated that the same annual precipitation with a different intra-annual temporal pattern indeed affected the runoff generation process (Tang et al., 2008). Since the runoff yield in excess of infiltration is the dominant runoff mechanism where precipitation intensity is the crucial driving force over most of the YRB (Jin et al., 2020), we focused on the impacts of the interannual precipitation and intra-annual monthly to annual precipitation ratio on the precipitation intensity.

Different precipitation intensities, including light, moderate, and heavy precipitation, are defined as daily precipitation amounts greater than 1, 10, and 25 mm, respectively in this region (Liu et al., 2018), and previous studies have proven that runoff was more sensitive to the total amount of heavy precipitation (P25) by analyzing a large number of in situ observation data (Liu and Gao, 2020). Therefore, the differences of interannual trends of P25 between S2 and S1 were calculated for each meteorological station to indicate the impact of interannual precipitation on the heavy precipitation, as demonstrated in Fig. 10a. The meteorological stations with an increasing trend in P25 driven by interannual precipitation change accounted for 69.7 %, with a maximum proportion of 80 % in the TDG–LM subregion, which caused the increase of the annual streamflow (Fig. 9).

Likewise, the impacts of the intra-annual monthly to annual precipitation ratio on the P25 were analyzed using the combination of S3 and S2, as shown in Fig. 10b. The meteorological stations with an increasing trend in P25 driven only by the intra-annual precipitation temporal patten change accounted for 58.9 %, hence the overall effect of the intra-annual temporal pattern change on the naturalized streamflow was positive during the study period. Spatially, increasing trends of P25 were observed in the majority of the stations within the source region (60 %) and LM–HYK (68 %), whereas the decreasing trend was dominant over the TNH–TDG and TDG–LM subregions, which led to the spatial heterogeneity of the effects of the intra-annual precipitation temporal patten change (Fig. 9).

To mitigate increasingly devastating ecological environment and soil erosion problems, the Grain for Green Project (GTGP), which targets the conversion of farmland into forests and grasslands (Jia et al., 2014; Liu et al., 2014), have been implemented over the upper and middle reaches of the YRB since 1998. According to the statistical data from local forestry authorities (Yao et al., 2011), afforestation in the Loess Plateau has mainly been implemented during 1998–2006, and the afforested areas across the plateau increased greatly from 3×104 km2 in 1998 to 5.9×104 km2 in 2006 (Fig. 11a). Between 1998 and 2006, artificially planted trees and shrubs increased rapidly, and the afforested areas of the Fenhe, Weihe, Beiluohe, Jinghe, Huangfuchuan, Yanhe, Kuyehe, and Wudinghe watersheds increased by 128 %, 113 %, 93 %, 82 %, 76 %, 66 %, 55 %, and 49 %, respectively, as shown in Fig. 11b. Previous studies on the Loess Plateau have suggested that compared with climate change, the tree and grass planting activities were the dominant driving factors for vegetation greening (Sun et al., 2015; Y. Zhang et al., 2016; Bai et al., 2018). In addition, natural rehabilitation without intensive interference activities, such as grazing prohibition, may play an important role in vegetation restoration in the Loess Plateau (Cao et al., 2011).

To explore the vegetation-type conversion caused by the GTGP, area percentage of the conversion of cropland into forest grass for each 0.1∘ grid cell during the study period was calculated using the GLC_FCS30. Figure 11c shows that massive vegetation-type conversion occurred in the TNH–HYK subregion with a maximum percentage of 50 %, which is partly proven by intense vegetation-type conversion detected using Landsat time series in the study of Wang et al. (2018). Figure 11d depicts the phenological characteristics of typical crop vegetation and the forest-grass vegetation. The LAI of forest-grass vegetation in the spring and autumn season is obviously higher than that of farm crops, whereas the LAI of crop vegetation in the summer season is slightly higher than that of forest-grass vegetation. Therefore, the massive vegetation-type conversion from cropland into forest-grass vegetation could significantly alter the vegetation phenology, which could lead to the interannual trend of the intra-annual monthly to annual LAI ratio increasing in spring and decreasing in summer (Fig. 6). Since phenology determines the start and end time of vegetation growth and is highly sensitive to climate change (Liang and Schwartz, 2009; Fu et al., 2019), climate warming has played an important role in advancing the spring phenology and delaying autumn phenology, and consequently extended the length of the vegetation growing period across the globe (Piao et al., 2019; Menzel et al., 2020), especially for the semiarid and semihumid regions of China (Wu et al., 2016; Chen et al., 2022). In addition, the variations in micro-topography from slope land into flat terrace significantly increase soil moisture (Bai et al., 2019), which could also inevitably alter interannual change and intra-annual temporal pattern of LAI.

Recent studies have increasingly focused on the effect of vegetation phenology and growth on runoff. It is found that earlier spring phenology and delayed autumn phenology promote a longer growing season and can increase the period for plant transpiration, potentially resulting in larger transpiration and might reduce the river runoff (Piao et al., 2019; Geng et al., 2020; Wu et al., 2012; Chen et al., 2022). These results were consistent with the negative effect of the intra-annual temporal pattern of LAI associated with the phenology change on runoff simulated by the VIC model, considering explicit vegetation dynamics in this study.

In general, previous studies evaluated the hydrological effects of vegetation change using the VIC model based on the multiyear average LAI and vegetation types during different periods (Xie et al., 2015; Yang et al., 2019; Zhai and Tao, 2021) as a result of the model configurations of VIC (Liang et al., 1994; Xie et al., 2007). However, due to the smoothing effect of averaging, the multiyear average LAI is unable to fully capture the vegetation continuous change, especially for the area with tremendous ecological restoration. Therefore, to explore the discrepancy in evaluating the hydrological effect of vegetation using VIC considering and without considering temporally explicit LAI change, we calculated the annual streamflow trend change by differencing the simulation of S1 with the simulation of the dynamic annual LAI observations while other variables varied under control conditions in S1. We then calculated the streamflow trend change using the combination of S1 and simulation where the annual LAI during 1982–1999 and 2000–2018 was fixed into the multiyear averages of the corresponding periods, respectively, while other variables varied in the same way as S1. Likewise, the annual streamflow trend changes simulated by continuous and noncontinuous change of the intra-annual temporal pattern of LAI were also calculated using the same way.

Figure 12 shows the comparison of the simulated annual streamflow trend using VIC considering and without considering continuous dynamics of interannual LAI and intra-annual temporal pattern of LAI in the TDG–LM (Fig. 12a and b) and LM–HYK (Fig. 12c and d) subregions. It is found that, compared with the simulation with multiyear average LAI changes, the impact of vegetation simulated by continuous LAI change increased by 42.9 % and 58.7 % for TDG–LM and LM–HYK, respectively, and the impact of vegetation simulated by the continuous intra-annual temporal pattern change was 3 times and 4.2 times of that simulated by noncontinuous inputs for TDG–LM and LM–HYK, respectively. These results were consistent with the reported attribution of runoff change in the upland Mediterranean basin where reductions in runoff were less intense when afforestation was not considered in the hydrological model (Buendia et al., 2015). It should be noted that although the change trends of simulated annual streamflow (Fig. 12b and d) are insignificant since the original interannual fluctuations were reserved in the scenario simulations, these trends become more significant when continuous LAI dynamics are considered in the VIC simulation, and the time series of the difference between simulated annual streamflow with VIC considering and without considering LAI dynamics show extremely significant change trend (P<0.001).

Previous studies focusing on this region at basin scale or regional scale have confirmed that massive vegetation greening has increased regional evapotranspiration through intense transpiration and canopy interception (Feng et al., 2016; Shao et al., 2019; Bai et al., 2018; Li et al., 2020), and caused a dried layer in the soil profile, interfering the vertical infiltration of soil water into the groundwater layer (Wang et al., 2011; Zhang et al., 2018), thus making negative impacts on the annual streamflow (Liang et al., 2015; Yang et al., 2019; Wang et al., 2021). Therefore, we further explore the impact of considering continuous LAI dynamic in VIC model on the simulations of total evapotranspiration and soil moisture of top-most layer in the middle reaches with significant vegetation increase. The discrepancies between VIC simulations with dynamic LAI and with fixed multiyear average LAI during 2000-2018 for annual total evapotranspiration and annual average soil moisture were calculated respectively, as illustrated in Fig. 13. The model using dynamic LAI tends to predict lower (higher) evapotranspiration and higher (lower) soil moisture than the model using static multiyear average LAI in the year when LAI was lower (higher), and the discrepancies were especially large for maximum annual anomaly of LAI, which is consistent with the findings of previous studies in the North America (Vivoni et al., 2008; Tang et al., 2012; Liu et al., 2018). This could explain the less intense reduction in runoff when continuous LAI increase was not considered in the hydrological simulation, as illustrated in Fig. 12.

To reduce sediment in the YRB, extensive water and soil conservation engineering measures including terraces and check dams were constructed over the Loess Plateau for mitigating soil erosion and intercepting sediment. According to the terrace proportion map (Cao et al., 2021) and statistical data about terrace areas of eight main tributaries (Liu et al., 2021), the built terrace was mainly distributed in TNH–HYK (Fig. 14a), and between 1979 and 2017 terrace areas of the Taohe, Huangshui, Qingshuihe, Beiluohe, Fenhe, Jinhe, Weihe subbasins and TDG–LM increased by 4.5, 4.9, 2.6, 3.0, 20.8, 10.4, 4.2, and 1.4 times, respectively (Fig. 14b), which indicated that change intensities of terrace areas in TNH–TDG and LM–HYK were greater than that in TDG–LM during the study period. Previous studies on the Loess Plateau have confirmed that due to the slope land changes into flat land, terraces can damage the continuity of the slope and prolong the infiltration time, resulting in poor hydrological connectivity and obvious runoff reduction (Tian et al., 2003; Bai et al., 2019). The study of Fu et al. (2020) also found that terraces play a critical role in reducing flood peak flow rate under extreme rainstorms.

In addition, check dams were increasingly built for blocking sediment from hillslopes into river channels, and the cumulative number of dams built above the Tongguan station during 1982–2015 was 3700 and 3010 for large-sized and medium-sized dams, respectively (Liu and Gao, 2020) (Fig. 14c). Although the check dam was originally designed to retain sediment, it still played a significant role in storing water for local crop irrigation, which has been captured by the significant increase of surface water area derived from the JRC product in the TNH–HYK subregion, as shown in Fig. 14d. It should be noted that greater change intensities in the terrace area and surface water area in TNH–TDG and LM–HYK compared with that in TDG–LM could probably explain the greater impact of residual factors on the streamflow reduction (Fig. 9) in these two subregions, which is consistent with the spatial pattern of impacts of residual factors on the evapotranspiration increase for the same subregions in the study of Wang et al. (2022).

For the source region where there are no significant changes in vegetation, in the last 3 decades, terraces and check dams reported that the degradation of permafrost attributed to climate warming and human activities could enhance active layer thickness above permafrost and decrease duration of seasonally frozen ground (Wu and Zhang, 2008; Cheng and Jin, 2013). This would have profound effects on the hydrology by altering soil surface infiltration capacity and soil hydraulic conductivity (Jin et al., 2009, 2011). When permafrost is thawed, it can be changed from an aquitard to an aquifer in some areas and talik channels can be formed or enlarged, which facilitate surface water infiltration, river runoff decrease, and groundwater recharge (Cheng and Jin, 2013). Cuo et al. (2013) found that it is highly possible that permafrost degradation played a role in diminishing river runoff. Meanwhile, increasing terrestrial water storage has also been confirmed in the study of Long et al. (2017).

The gridded forcing data may introduce uncertainties in the simulations because these climate data are interpolated based on limited field observations. It would be better to merge high-accuracy microwave precipitation products and reanalysis data in the future. The GLASS LAI data were only used here; although differences exist between different LAI products, these LAI products are generally consistent in the spatiotemporal changes across China (Piao et al., 2015; Zhu et al., 2016). Hence this would probably not change the general conclusions (Zhai et al., 2021).

It is difficult to accurately acquire naturalized streamflow due to some uncertainties regarding data of human water use, especially from irrigation, which could explain the NSE lower than 0.5 in the validation period (Table 4) in TNH–TDG and LM–HYK where there are large irrigated areas. In addition, all grid cells of the subregions were characterized with a constant parameter dataset based on an idealized assumption. Hence further calibration should be conducted in more subbasins by collecting high-quality naturalized hydrological data and using a hydrological model considering human water use to mitigate uncertainties of model parameters.

The scenario simulation method would split the link and interaction between climate change and underlying surface change (Wu et al., 2017), which inevitably introduces a certain bias in quantifying the variation in streamflow induced by interannual and intra-annual changes of climate variables and vegetation. Even though the interactive effect was calculated by differencing the sum of variations in the streamflow that were induced by climate and vegetation change and those induced by the coefficient in this study, this simplified method still cannot represent complicated feedback and response of climate and underlying surface change. The LAI increase is always associated with land cover change, hence the vegetation's hydrological effect was considered as the total impact from LAI and land cover changes in this study. This inevitably involves the impacts of non-vegetated land cover conversion (e.g., urbanization). Nevertheless, this land cover change type only accounts for a very small proportion of the YRB.

Since the lack of water consumption data of coal mining and the effects of glaciers melting and permafrost degradation on the runoff generation were not considered during the VIC simulation in this study, the impacts from coal mining, glacier, and permafrost in analyzing the relationship between non-vegetation underlying surface change and river runoff were not further clarified. In addition, only one model was applied here, and water and soil conservation engineering measures were not considered in the model. The conclusions of this study should be proven in further studies by combining the statistical model, lumped model, distributed model, and machine learning model.