The TRSR (30–36∘ N, 90–104∘ E) is located in the hinterland of the QTP, with an altitude ranging from 2677 to 6575 m (Fig. 1). It contains the headwaters of the Yangtze, Yellow, and Lancang rivers, covering an area of approximately 36.1×104 km2 (Ji et al., 2020). The TRSR's climate is a typical plateau continental climate, characterized by low temperatures, strong radiation, and no obvious distinction between four seasons (Tong et al., 2010). During 1981–2010, the annual mean precipitation was 593 mm, and the annual mean temperature was 1.9 ∘C on the TRSR (Luo et al., 2017), which both generally decreased from southeast to northwest (Deng and Zhang, 2018). Approximately 75 % of total annual precipitation occurs from June to September, and the temperature is always highest during this period (Fig. S1 in the Supplement). The average annual runoff within the TRSR is about 47.5×109 m3. Glaciers are widely distributed in the TRSR, with about 1700 glaciers covering an area of about 2300 km2 in all.

Three kinds of data were included in this study, namely, meteorological forcing data, land surface characteristic data, and runoff data. The meteorological forcing data – including daily precipitation; wind speed; and maximum, minimum, and mean temperatures from 1983 to 2018 – were obtained from the China Meteorological Administration (CMA) (http://data.cma.cn, last access: 3 April 2023). Most weather stations are located in the south and southeast of the TRSR, with very few in the central and western regions (Fig. 1). For the Yangtze River source region, the interpolated precipitation based on CMA rain gauges (hereinafter called CMA precipitation) cannot capture the spatial details well (X. Liu et al., 2017; Xue et al., 2013). Therefore, in this subregion, we also used Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks – Climate Data Record (PERSIANN-CDR) (https://www.ncei.noaa.gov/products/climate-data-records/precipitation-persiann, last access: 3 April 2023), which has the advantages of high spatial (0.25∘×0.25∘) and temporal (daily) resolution, long time span (more than 35 years since 1983, and the data range is constantly updated), and more complete coverage (Ashouri et al., 2015). Previous studies have also confirmed that PERSIANN-CDR is a high-quality precipitation data set applicable to the source region of the Yangtze River (X. Liu et al., 2017; Y. Wang et al., 2021).

The soil texture data came from the global 5 arcmin data set of the Food and Agriculture Organization of the United Nations (FAO), and the vegetation types were provided by the global 1 km land cover classification database produced by the University of Maryland (http://prettymap.mooncoder.com/maps/metadata/data.html, last access: 3 April 2023). The Shuttle Radar Topography Mission (SRTM) digital elevation data set with a resolution of 90 m was obtained from the Geospatial Data Cloud (http://www.gscloud.cn/search, last access: 3 April 2023). Glacier area data were from TPG1976 generated by Ye et al. (2017). This data set was specially compiled for the QTP based on Landsat satellite images from the mid-1970s, among which most images were acquired from 1976, and SRTM digital elevation models (DEM v4.1) and Google Earth imagery. We assumed that the glacier area around the source of the Yellow River was 0 km2 due to its proportion of the whole area and contribution to total runoff being minor (Y. Wang et al., 2021; Zhang et al., 2013).

Monthly runoff observation data were from the corresponding hydrological stations. The details of the time ranges of the available data are shown in Table 1. It should be noted that there was no observed runoff from November to April at Tuotuohe station and that, in this study, we treated the runoff in these months as 0 because the runoff was negligible (Ahmed et al., 2020; Luo et al., 2019).

Figure 2 shows the simulated and observed monthly runoff of the seven stations and summarizes the model's performance. In general, the model achieved reasonably satisfactory results, with the NSE exceeding 0.68 at all stations. But there still existed a certain degree of discrepancy between simulations and observations in some years for some stations, such as when the simulation slightly overestimated runoff peaks at Zhimenda station in the validation period, while it underestimated Changdu and Xialaxiu stations' lowest runoff in both the calibration and validation periods. Uncertainty arising from model parameters and inferred runoff may have caused these discrepancies, but the most likely and important reason may be the over-/underestimation of precipitation forcing (Miao et al., 2022; Su et al., 2016).

Accurate precipitation input is a prerequisite for obtaining reasonable model parameters and simulation results (Zhang et al., 2013; Chen et al., 2017). Figure 3 compares the simulation results for monthly runoff driven by CMA precipitation and PERSIANN-CDR precipitation at Tuotuohe and Zhimenda stations in the source area of the Yangtze River, where meteorological stations are extremely sparse (Fig. 1). The CMA-precipitation-driven model exhibited poor performance even after parameter optimization, except for at Zhimenda station during the calibration period. In contrast, taking NSE as an example, and in comparison with the result obtained using CMA precipitation, the NSE of Tuotuohe station increased from 0.31 to 0.86 when PERSIANN-CDR precipitation was used during the validation period. Such a large improvement indicates that the PERSIANN-CDR precipitation, which describes the spatial variation of precipitation more accurately than the CMA precipitation obtained through very limited rain-gauge interpolation in the source area of the Yangtze River (X. Liu et al., 2017; Bai and Liu, 2018), can be used as input for the model to generate more accurate runoff. This sheds light on the importance of precipitation to hydrological research and the prospect of using PERSIANN-CDR precipitation products in alpine regions with sparse meteorological stations (X. Liu et al., 2017). Because the model with CMA precipitation as input performed well in the source regions of the Yellow River and the Lancang River, there was no experiment using PERSIANN-CDR precipitation in these two headwater subregions.

Figure 4 shows the contributions of rainfall, snowmelt, and glacier runoff to the total annual runoff at the seven stations for 1984–2018. We found that rainfall runoff was the main component of runoff, accounting for 82 %–92 %, whereas snowmelt and glacier runoff both accounted for less than 10 %, due to low temperatures and small glaciers (Table 1). Our current results are similar to the previous finding that snowmelt runoff and glacier runoff both make up a small proportion of the total runoff in the TRSR, but note that the specific values of their contributions to runoff differed among these studies. Y. Wang et al. (2021) found that during 1984–2015, snowmelt runoff contributed 15 % to the total runoff at Zhimenda station. This estimate is higher than the 8.9 % contribution in 1984–2018 estimated by our current work and the 7 % contribution in 2003–2014 estimated by Han et al. (2019). The discrepancies among these results may be attributed to differences specific to the research periods, but various choices of forcing input data, model parameters, and definitions of snowmelt runoff and glacier runoff should be more important factors (Y. Wang et al., 2021; Zhao et al., 2019; Sun and Su, 2020).

Figure 5 shows the monthly variation in each runoff component. The regime of rainfall runoff of all stations was highly consistent with that of total runoff, which once again confirms the leading role of rainfall in total runoff. Snowmelt runoff mainly occurred from April to June due to the melting of winter snowpack and spring snowfall, with contributions ranging from 20 % to 60 % (Fig. S2). As a result of higher altitude (Fig. 1) and lower temperature (Deng and Zhang, 2018), the snowmelt runoff peak at Tuotuohe and Zhimenda stations in the source area of the Yangtze River emerged in June, while this peak occurred at the other stations in May. Glacier runoff was mainly concentrated in July and August, corresponding to the higher temperatures during this period (Fig. S1). However, due to the fact that glaciers made up only a small proportion of the total area (Table 1), the monthly contribution of glacier runoff was below 25 %.

Compared with the reference period (1984–2003), the mean annual runoff in all sub-basins of the TRSR during the changed period (2004–) increased (Fig. 6a), especially in the source area of the Yangtze River, with the catchments monitored by Zhimenda and Tuotuohe stations increasing by 31 % and 51 %, respectively. Figure 7 presents the absolute impacts and relative contributions of four influencing factors (precipitation, temperature, interactions of climatic variables, and human activity) on the annual runoff increase in different sub-basins. It is clear that precipitation was strongly and positively correlated with the runoff, with relative contributions of 38 %–71 %, in agreement with the findings of Wu et al. (2018), who reported that increasing precipitation can directly increase runoff. Because it has increased the most significantly in the source region of the Yangtze River during recent decades (Fig. 6b), precipitation had a greater impact on runoff variation there than it did in the other two headwater subregions.

Although the runoff in catchments monitored by Tuotuohe and Xiangda stations exhibited a minor increase (≤5 %) when the temperature rose, warming temperature had an overall effect of reducing runoff (by 4 %–25 %) because the limited increase in meltwater was largely offset by the enhancement of evapotranspiration and the degradation of frozen soil (Zhao et al., 2019). Given the spatiotemporal changes in climate variables and their complex interaction, the interactive effects of climate variables on runoff change were spatially non-uniform within the TRSR. Specifically, the absolute impacts of this factor on runoff change varied from -11 to 20 mma-1, and the relative contribution ranged from 0 % to 26 %. Human activity interference in the catchment monitored by Xialaxiu station far exceeded that in other catchments. It explains the 49 % increase in runoff, which can be attributed to the significant degradation of grassland in this region during the changed period (Zeng et al., 2021; Zhang et al., 2021). In the context of increasing water demand related to agriculture and industry (Zhai et al., 2021) and the ecological protection policy proposed in the 21st century (D. Liu et al., 2017), human activities have consistently reduced runoff in the other catchments, except in the catchment monitored by Changdu station.

Overall, the impacts of precipitation, temperature, the interactions of climatic variables, and human activities on runoff change present differences among different sub-basins. Climate change, which integrates precipitation- and temperature-induced change with that due to the interactions of climatic variables, accounted for over 75 % of the change in runoff for all basins except the catchment monitored by Xialaxiu station, where the contribution of climate change was 51 %. Therefore, in the TRSR, the dominant factor influencing runoff variation was climate change, especially precipitation.

Figure 8 shows projected percent changes of total runoff and each runoff component under four climate change combination scenarios with respect to the period from 1984 to 2018. The total runoff at all stations was expected to increase the most under scenario S3 (P+20 %, T-1 ∘C), with an increase of 29 %–80 %, followed by scenario S4 (P+20 %, T+1 ∘C), with an increase of 25 %–41 %, whereas under scenarios S1 (P-20 %, T-1 ∘C) and S2 (P-20 %, T+1 ∘C), respective decreases of 24 %–34 % and 21 %–56 % were predicted. These results indicate that the total runoff was mainly affected by precipitation, but the magnitude of response could strengthen or weaken with changing temperature, which affects not only evapotranspiration but also meltwater (Su et al., 2016). Spatially, as a consequence of neglected glacier runoff, total runoff changes at Tangnaihai and Maqu stations in the source region of the Yellow River were maximal under scenario S3 in comparison with other stations. We further found that the total runoff showed a higher sensitivity to precipitation increase than to precipitation decrease when the temperature remained constant relative to the period of 1984–2018 (Fig. S3). Taking Zhimenda station as an example, the total runoff would increase by 50 % if precipitation increased by 20 % but decrease by about 41 % if precipitation decreased by 20 %. This may be explained by the runoff generation process, just as Spencer et al. (2019) reported that a continuous multi-year pattern of lower- or higher-than-average precipitation can reduce or fill basin storage and affect runoff responses. The different behaviors of total runoff when only temperature changes are shown in Fig. S4. For the catchment monitored by Xiangda station, total runoff was basically unchanged when the temperature changed. This can be explained by the fact that temperature has a similar degree of influence on evapotranspiration and meltwater within this region, while evapotranspiration plays a stronger role in other sub-basins. The pattern of change in rainfall runoff was essentially consistent with that of total runoff, which is closely associated with precipitation although temperature change will affect evapotranspiration.

As climatic variables changed, snowmelt runoff tended to have a larger degree of variation (-76 % to 203 %) than other runoff components did, indicating that it is more susceptible to climate change. Despite its similarity to rainfall runoff and total runoff, snowmelt runoff also increased most obviously in scenario S3 (P+20 %, T-1 ∘C). The major reasons for this were the direct increase in precipitation and the increased proportion of precipitation falling as snow at the lower temperature (Chandel and Ghosh, 2021). Snowmelt runoff varied little in the wetter and warmer scenario (scenario S4) relative to 1984–2018 because increasing precipitation was basically entirely compensated for by the decreased solid-to-liquid precipitation ratio. The most significant change in snowmelt runoff was observed at Tuotuohe station and can be explained by this station's relatively high elevation, which amplifies the importance of snowmelt (Fig. 1). Glacier runoff was highly dependent on temperature change whether precipitation increased or decreased. Consistent with the small area proportion of glaciers in each sub-basin (Table 1), the variation in glacier runoff did not show obvious spatial heterogeneity.

Figure 9 presents the projected seasonal change percentage of total runoff under different scenarios. As shown in Figs. 9 and S5, on the seasonal scale, similar to the annual scale, the change pattern of total runoff was similar to that of rainfall runoff, further confirming the importance of rainfall runoff in the TRSR. Additionally, due to over 80 % of precipitation occurring in summer (July–September) and autumn (October–December) (Fig. S1), more obvious total runoff changes were projected during this period. Although temperature was poorly and negatively correlated with the total runoff for most stations at the seasonal scale, the total runoff at Tuotuohe station increased by 10 % with a temperature increase of 1 ∘C in spring (April–June) as the result of the release of more meltwater, which could possibly advance the peak flow of snowmelt (Shen et al., 2018). Su et al. (2016) have also reported that in the source region of the Yangtze River, an apparent earlier melt may happen in April under a warming climate. Due to the neglected glacier runoff and the more obvious warming trend compared with the other two source areas in the past few decades (Yi et al., 2011; Xie et al., 2004), the total runoff of the Yellow River source displayed a more strongly significant change (-18 % to 25 %) in summer when the temperature changed relative to 1984–2018 (Fig. 9b). Generally speaking, the snowmelt runoff has a second small peak because of the melting of fresh snowfall in autumn (Zhang et al., 2013), as well as the high temporal and spatial variability of precipitation in the source region of the Yangtze River (X. Liu et al., 2017), leading to Tuotuohe station's snowmelt runoff showing the highest sensitivity to climate change in autumn (Fig. S6). The catchment monitored by Xiangda station has more glacier coverage (Table 1), which may be why its glacier runoff had a greater response to temperature in winter in comparison with the glacier runoff in the other basins (Fig. S7).

In general, precipitation plays a greater role in total runoff and rainfall runoff, while glacier runoff is dominated by temperature at annual and seasonal scales. Snowmelt runoff shows a different behavior. Annually, it is affected by both precipitation and temperature. In spring and summer, it is more related to precipitation, in winter it is more related to temperature, and in autumn it is more related to the combined effects of temperature and precipitation.

This work provides a systematic understanding of hydrological processes in the TRSR. The results are encouraging, but some uncertainties are still worthy of further analysis in future research. Due to the lack of observed runoff in Tuotuohe station from November to April, we treated the runoff as 0. Although previous studies have shown that the runoff in these months is negligible (Ahmed et al., 2020; Luo et al., 2019), it inevitably introduced some uncertainty because the missing data would affect the model evaluation metrics, thus affecting the optimization results of model parameters. In addition to this, we simulated glacier runoff using a simple degree-day factor algorithm, which is highly sensitive to its DDF parameter (Chen et al., 2017; Zhang et al., 2013; Zhao et al., 2019). Given the lack of observed glacier information, we did not calibrate DDF but directly adopted the constant values at basin scale from existing studies that did not take into account the high spatiotemporal heterogeneities of DDF (Zhang et al., 2006). Although the final simulation result was satisfactory, the practice of treating DDF as a uniform value still introduced a certain degree of bias to the results. In fact, due to the very limited glacier meltwater contribution (Fig. 4), the uncertainty caused by glacier runoff is far less than that stemming from precipitation forcing data (Zhao et al., 2019). Many previous studies have reported that precipitation is the key factor limiting hydrological simulation performance in alpine regions with sparse meteorological stations (X. Liu et al., 2017). Precipitation estimation error will bring significant uncertainty to runoff simulations (Sun and Su, 2020). As shown in Fig. 3, the model's performance was greatly improved by replacing CMA precipitation with PERSIANN-CDR data in the source region of the Yangtze River. This finding inspires us to hope that satellite-based precipitation products may be suitable alternatives for hydrological studies in alpine regions and that we can use several sets of precipitation products to reduce the uncertainty of simulation as much as possible in the future.

Some uncertainties may also exist in assessing the influence of climate change and human activities on changes in runoff. Hydrological model simulation splits the relationship and interaction between climate change and human activity, which inevitably introduces a certain bias due to the complicated feedback and response relationship between them (Ji and Yuan, 2018a; Shi et al., 2022). And our study focused only on the effects of long-term mean annual precipitation and temperature change on runoff without considering the possible impact of the intra-annual fluctuations of these factors. Therefore, additional work should be carried out to examine the causes of runoff change in more detail.