1 Introduction 1 引言

The hydrological cycle is a key component of the Earth's climate system (Wu et al., 2013) and is one among the different cycles operating in nature, such as the carbon cycle, the nitrogen cycle, and other biogeochemical cycles (Jain & Singh, 2017). It is recognized worldwide that water plays a major role in the socio-economic development (e.g., energy production, agriculture, domestic and industrial water supply) of a country and is a critical component of the global as well as the regional environment (Cosgrove & Loucks, 2015; Miao & Ni, 2009). However, the available freshwater resources are under great pressure throughout the world due to rapidly increasing population (Islam & Karim, 2020), extending irrigated agriculture (Fischer et al., 2007), developing industry (Flörke et al., 2013), and economic and technological development (V. Singh et al., 2014), which can lead to a critical water resource shortage (Xuan et al., 2018). Moreover, the hydrological cycles of different regions have been intensified throughout the world due to the integrated consequences of global warming, changing climate, and human interference (Cohen et al., 2014; Madakumbura et al., 2019).
水文循环是地球气候系统的关键组成部分（Wu 等人，2013 年），也是碳循环、氮循环和其他生物地球化学循环等在自然界中运行的不同循环之一（Jain 和 Singh，2017 年）。世界公认，水在一个国家的社会经济发展（如能源生产、农业、家庭和工业供水）中发挥着重要作用，是全球和区域环境的重要组成部分（Cosgrove & Loucks，2015；Miao & Ni，2009）。然而，由于人口的快速增长（Islam & Karim，2020 年）、灌溉农业的扩大（Fischer 等人，2007 年）、工业的发展（Flörke 等人，2013 年）以及经济和技术的发展（V. Singh 等人，2014 年），全世界可用的淡水资源正面临着巨大的压力，这可能会导致严重的水资源短缺（Xuan 等人，2018 年）。此外，由于全球变暖、气候变迁和人为干扰的综合后果，世界各地不同区域的水文循环已经加剧（Cohen 等人，2014 年；Madakumbura 等人，2019 年）。

Anthropogenic climate change is considered one of the biggest threats to freshwater resources in the twenty-first century (Capon et al., 2021) because it can disturb the stability and availability of water resources at local, regional, and national scales. According to the Intergovernmental Panel on Climate Change (IPCC), the global air temperature has increased by 1.1°C since 1880, at a rate of 0.08°C/10 year. The rate of warming is even more than twice (0.18°C/10 year) since 1980 (IPCC, 2022a; NOAA-NCEI, 2023). The situation might be worse as the global temperature is projected to increase by 2.1–3.5°C under SSP2-4.5 and 3.3–5.7°C under SSP5-8.5 by the end of the twenty-first century relative to the pre-industrial era (IPCC, 2021). Climate change impacts have been reported across every ecosystem on the planet (Scheffers et al., 2016). For example, climate change has induced substantial damages and even increasingly irreversible losses in the terrestrial, coastal, marine, and freshwater ecosystems. In addition, frequent and intense extreme events due to human-induced climate change have reduced food and water security, causing hindrances in meeting sustainable development (IPCC, 2022b). Among them, freshwater resources (e.g., rivers, lakes, and wetlands) are even highly vulnerable and have the potential to be strongly impacted by climate change threats (IPCC, 2008). For example, renewable surface water and groundwater resources are projected to reduce significantly in most dry-subtropical regions, and the frequency of meteorological droughts (less rainfall) and agricultural droughts (less soil moisture [SM]) is likely to increase in dry regions due to climate change. Moreover, climate change negatively impacts freshwater resources by altering the streamflow regimes and water quality (IPCC, 2014).
人为气候变化被认为是二十一世纪淡水资源面临的最大威胁之一（Capon et al.根据政府间气候变化专门委员会（IPCC）的数据，自 1880 年以来，全球气温上升了 1.1°C，上升速度为 0.08°C/10。自 1980 年以来，升温速度甚至超过了两倍（0.18°C/10 年）（IPCC，2022a；NOAA-NCEI，2023）。预计到 21 世纪末，全球气温与前工业化时代相比，在 SSP2-4.5 条件下将上升 2.1-3.5°C，在 SSP5-8.5 条件下将上升 3.3-5.7°C，情况可能会更糟（IPCC，2021 年）。据报道，气候变化对地球上的每个生态系统都产生了影响（Scheffers 等人，2016 年）。例如，气候变化对陆地、沿海、海洋和淡水生态系统造成了巨大破坏，甚至造成了越来越多不可逆转的损失。此外，人类活动引起的气候变化导致极端事件频发且强度大，降低了粮食和水安全，阻碍了可持续发展（IPCC，2022b）。其中，淡水资源（如河流、湖泊和湿地）甚至非常脆弱，有可能受到气候变化威胁的强烈影响（IPCC，2008 年）。例如，在大多数亚热带干旱地区，可再生地表水和地下水资源预计将大幅减少，而且由于气候变化，干旱地区气象干旱（降雨量减少）和农业干旱（土壤水分减少[SM]）的频率可能会增加。 此外，气候变化还会改变水流机制和水质，从而对淡水资源产生负面影响（IPCC，2014 年）。

However, the impacts of climate change on water resources vary from region to region (J. Chu et al., 2010; Mahmood & Jia, 2019). For example, the Three-River Headwater Region (TRHR), the source of three giant rivers of Asia (i.e., the Yangtze, Yellow, and Lancang), has experienced significant hydrological changes because of climate change in the past few decades (T. Su et al., 2023). It has a very fragile ecosystem and has extremely sensitive water resources to changing climate (L. Zhang et al., 2017). It is located in an arid-semiarid zone of the Qinghai Tibetan Plateau (QTP), which is also known as China's water tower (Z. Wang et al., 2010) and represents an important ecological barrier in China (Jiang & Zhang, 2016). It approximately produces an annual discharge of 40 BCM and contributes a considerable amount of water to the downstream regions (Mahmood et al., 2020), which are home to more than 500 million people (Varis et al., 2014). Therefore, it plays a critical role in food security and eco-environmental protection in China and even Southeast Asia (T. Su et al., 2023). However, many studies such as Bei et al. (2019), Deng et al. (2019), Liang et al. (2013), Lu et al. (2018), Shi et al. (2016), You et al. (2014), Yuan et al. (2018), and Y. Zhang et al. (2020), showed noticeable changes in the climate of the region, especially an increased temperature, ranging from 0.28 to 0.54°C/10 year, and precipitation, ranging from 5 to 20 mm/10 year, since the 1950s and 1960s. The region has even experienced twice the rate of the global average temperature (0.18°C/10 year since 1980) (IPCC, 2022a; T. Su et al., 2023). The changing climate along with profound human interventions, specifically the projects related to the protection and restoration of the ecological environment, have disordered the hydrological processes in the region (T. Su et al., 2023). For example, due to increased temperature, actual evapotranspiration has been increasing at a rate of 4–15 mm/10 year since the 1980s (Bei et al., 2019; Yuan et al., 2018), while up to 60 mm/10 year after 2003 (X. Li et al., 2019). In addition, regional warming along with increased precipitation has caused rapid glaciers retreat, permafrost degradation, reduction in the active layer of frozen soil, altered streamflow regimes, increased extreme events (e.g., precipitation storms, heat waves, and flooding), biodiversity loss, and ecosystem disturbance (Mahmood et al., 2020; R. Wang et al., 2023; Q. Zhao et al., 2019). Trend analysis conducted by different studies such as H. Chu et al. (2019), Jiang et al. (2016), and T. Mao et al. (2016) in the TRHR have shown significant increased streamflow in the Headwater of the Yangtze River (HYaR) and Headwater of the Lancang River (HLaR), while decreased streamflow in the headwater of the Yellow River (HYeR). These alterations in streamflow regimes are mainly attributed to climate change, about 75%–90%, instead of human interventions (Jiang et al., 2017; T. Su et al., 2023).
然而，气候变化对水资源的影响因地区而异（J. Chu 等人，2010 年；Mahmood & Jia，2019 年）。例如，三江源地区（TRHR）是亚洲三条大河（即长江、黄河和澜沧江）的源头，在过去几十年中，由于气候变化，该地区的水文发生了显著变化（T. Su 等人，2023 年）。它的生态系统非常脆弱，水资源对气候变化极为敏感（L. Zhang 等人，2017 年）。它位于青藏高原（QTP）的干旱半干旱地带，也被称为中国的水塔（Z. Wang 等人，2010 年），是中国重要的生态屏障（Jiang & Zhang，2016 年）。它的年排水量约为 400 亿立方米，为下游地区提供了大量的水源（Mahmood 等人，2020 年），而这些地区是 5 亿多人口的家园（Varis 等人，2014 年）。因此，它在中国乃至东南亚的粮食安全和生态环境保护中发挥着至关重要的作用（T. Su 等人，2023 年）。然而，许多研究，如 Bei 等人（2019 年）、Deng 等人（2019 年）、Liang 等人（2013 年）、Lu 等人（2018 年）、Shi 等人（2016 年）、You 等人（2014 年）、Yuan 等人（2018 年）和 Y. Zhang 等人（2020 年）显示，自 20 世纪 50 年代和 60 年代以来，该地区的气候发生了明显的变化，尤其是气温增加了 0.28 至 0.54°C/10，降水增加了 5 至 20 毫米/10。该地区的气温甚至是全球平均气温的两倍（自 1980 年以来为 0.18°C/10）（IPCC，2022a；T. Su 等人，2023 年）。 不断变化的气候以及人类的大量干预，特别是与保护和恢复生态环境相关的项目，使该地区的水文过程发生紊乱（T. Su 等人，2023 年）。例如，由于气温升高，自 20 世纪 80 年代以来，实际蒸散量以每年 4-15 毫米/10 年的速度增加（Bei 等人，2019 年；Yuan 等人，2018 年），而 2003 年之后则高达 60 毫米/10 年（X. Li 等人，2019 年）。此外，随着降水量的增加，区域变暖导致冰川快速后退、永久冻土退化、冻土活性层减少、溪流机制改变、极端事件（如降水风暴、热浪和洪水）增加、生物多样性丧失和生态系统紊乱（Mahmood 等，2020 年；R. Wang 等，2023 年；Q. Zhao 等，2019 年）。不同研究（如 H. Chu 等人，2019 年）、Jiang 等人，2016 年）和 T. Mao 等人，2016 年）对 TRHR 进行的趋势分析表明，长江源头（HYaR）和澜沧江源头（HLaR）的流量显著增加，而黄河源头（HYeR）的流量减少。河水流量变化的主要原因是气候变化，约占 75%-90%，而非人为干预（Jiang 等人，2017 年；T. Su 等人，2023 年）。

These hydro-climatic situations in the TRHR might get worse in the future, especially, after the mid-twenty-first century and under SSP5-8.5 as different studies such as Hu et al. (2022), Ji et al. (2020), L. Liu et al. (2011), Lu et al. (2018), Lutz et al. (2014), F. Su et al. (2016), T. Wang et al. (2022), Y. Zhang et al. (2015), and Q. Zhao et al. (2019) have already shown some serious alterations in hydrological regimes, using the outputs of Global Climate Models (GCMs). For example, T. Wang et al. (2022) projected temperature, precipitation, evapotranspiration, and runoff using the bias-corrected ensemble mean of 5 GCMs under 3 SSPs (i.e., 1, 2, and 5) for 2021–2070 in the upper Yangtze River (above Yichang) and showed increased runoff above Zhimenda while overall decrease, though significantly increased precipitation in future. Hu et al. (2022) used a similar methodology (bias-corrected ensemble of 8 GCMs) under different representative concentration pathways (RCPs) to explore the streamflow responses to climate change in the HYaR and showed a 15%–20% decreased streamflow. They concluded decrease in streamflow was mainly due to highly projected increased evapotranspiration (30%–54%). Similarly, Ji et al. (2020) used bias-corrected outputs of 6 GCMs to investigate the extreme hydrological events in the TRHR under global warming levels of 1.5, 2.0, and 3.0°C and found a significant increase in dry-extremes over the HYeL while wet-extreme over the HYaR. However, according to Immerzeel et al. (2010), the runoff in the upper Yangtze basin (5%) will decrease while will increase in the upper Yellow (9%) under A1B. Similarly, Q. Zhang et al. (2017) also explored increased streamflow in the HYeR, especially under RCP8.5. These contradicting results show that there is still a need to investigate future changes in hydrological regimes under changing climate using recently released the GCM's climate data by the Coupled Model Intercomparison Project Phase 6 (CMIP6) because few studies are reported using recently released SSP scenarios in the TRHR. Moreover, the previous studies concentrated mainly on precipitation, streamflow, and evapotranspiration of the water balance components, though other hydrological components such as surface flow (direct flow), baseflow, and SM contents, snowmelt flow, and terrestrial water storage (TWS) are equally important in such important region, which is composed of lakes, rivers, wetlands, glaciers, permafrost, seasonal frozen soil, and snow-covered mountains (Tong et al., 2014).
(2018)、Lutz 等人(2014)、F. Su 等人(2016)、T. Wang 等人(2022)、Y. Zhang 等人(2015)和 Q. Zhao 等人(2019)已经利用全球气候模式(GCMs)的输出结果表明水文系统发生了一些严重变化。例如，T. Wang 等人（2022 年）在 3 个 SSPs（即 1、2 和 5）下使用 5 个 GCMs 的偏差校正集合平均值预测了长江上游（宜昌以上）2021-2070 年的温度、降水、蒸散和径流，结果表明，虽然未来降水量显著增加，但志门达以上地区的径流量增加，而总体降水量减少。Hu 等人（2022 年）使用类似方法（8 个 GCM 的偏差校正集合），在不同的代表性浓度路径（RCPs）下，探讨了 HYaR 中的溪流对气候变化的响应，结果显示溪流减少了 15%-20%。他们得出的结论是，流量减少的主要原因是高度预测的蒸散量增加（30%-54%）。同样，Ji 等人（2020 年）使用 6 个全球环流模型的偏差校正输出结果，研究了 TRHR 在全球变暖 1.5、2.0 和 3.0 摄氏度的情况下的极端水文事件，发现 HYeL 的极端干流显著增加，而 HYaR 的极端湿流显著增加。然而，根据 Immerzeel 等人（2010 年）的研究，在 A1B 条件下，长江上游流域的径流将减少（5%），而黄河上游流域的径流将增加（9%）。同样，Q. Zhang 等人（2017 年）也探讨了多年平均径流量的增加，尤其是在 RCP8.5 条件下。 这些相互矛盾的结果表明，仍然需要利用耦合模式相互比较项目第 6 阶段（CMIP6）最近发布的 GCM 气候数据来研究气候变化下水文系统的未来变化，因为在 TRHR 中利用最近发布的 SSP 情景的研究报告很少。此外，以往的研究主要集中在水量平衡成分中的降水、溪流和蒸发蒸腾，而其他水文成分，如地表流（直接流）、基流和 SM 含量、融雪流和陆地蓄水（TWS），在由湖泊、河流、湿地、冰川、永久冻土、季节性冻土和雪山组成的这一重要区域同样重要（Tong 等，2014 年）。

The common thing we noticed in the previous studies is applying bias-corrected outputs of GCMs because of its simplicity, and to the best of our knowledge, no studies were found applying a sophisticated statistical downscaling technique such as statistical downscaling model (SDSM), where the simulated climate data is first evaluated against the historical observations before using them for the future projections. The SDSM technique is a combination of multiple linear regression (MLR) and stochastic generator and is the most common and sophisticated method (Huang et al., 2011; Mahmood & Babel, 2013) for downscaling climate variables such as temperature and precipitation (Fan et al., 2021; Yang et al., 2017). Different comparative studies have shown that this technique performed relatively better than other methods such as bias correction, as in Campozano et al. (2016), Hernanz et al. (2022), and W. Liu et al. (2013). However, screening (selection) of large-scale variables (predictors), which is central to all statistical downscaling techniques (Wilby et al., 2002), is subjective, time-consuming, and also requires expertise to select some suitable predictors by considering the effect of multi-collinearity (Hammami et al., 2012; Mahmood & Babel, 2013). In this study, we developed a quantitative screening method, which not only minimizes the collinearity but significantly reduces the number of predictors and also ranks them. This is the modified form of the screening method used by Mahmood and Babel (2013). They used a reduction formula of partial correlation (PC) and correlation coefficient between predictand and predictor (CCPP). The main problem in this formula is it, sometimes, selects the predictors with lower CCPP instead of predictors of high CCPP when the difference between PC and CCPP is minimum without considering the CCPP. To overcome this problem, instead of using a reduction formula, we used a product formula of PC and CCPP, along with P-value and backward regression. This showed significant improvement when compared with the results obtained by the stepwise regression method, which is considered a classical method for the screening of predictors (Hammami et al., 2012). Since a large number of predictors limits the number of GCMs, this method can help in including more GCMs and scenarios in the analysis, which is recommended for impact studies.
在以往的研究中，我们注意到的一个共同点是，由于其简单性，我们采用的是偏差校正后的 GCMs 输出结果，而据我们所知，没有发现任何研究采用了复杂的统计降尺度技术，如统计降尺度模型 (SDSM)，即在将模拟气候数据用于未来预测之前，首先根据历史观测数据对其进行评估。SDSM 技术是多元线性回归（MLR）和随机生成器的结合，是对温度和降水等气候变量进行降尺度的最常见、最复杂的方法（Huang 等人，2011 年；Mahmood & Babel，2013 年）（Fan 等人，2021 年；Yang 等人，2017 年）。不同的比较研究表明，该技术的性能相对优于其他方法，如偏差校正，如 Campozano 等人（2016 年）、Hernanz 等人（2022 年）和 W. Liu 等人（2013 年）。然而，大规模变量（预测因子）的筛选（选择）是所有统计降尺度技术的核心（Wilby 等人，2002 年），它是主观的、耗时的，还需要专业知识，通过考虑多重共线性的影响来选择一些合适的预测因子（Hammami 等人，2012 年；Mahmood & Babel，2013 年）。在本研究中，我们开发了一种定量筛选方法，它不仅能最大限度地减少共线性，还能显著减少预测因子的数量并对其进行排序。这是 Mahmood 和 Babel（2013 年）使用的筛选方法的改进形式。他们使用了部分相关性（PC）和预测因子与预测因子之间的相关系数（CCPP）的还原公式。 该公式的主要问题在于，当 PC 与 CCPP 的差值最小时，它有时会选择 CCPP 值较低的预测因子，而不考虑 CCPP 值较高的预测因子。为了克服这个问题，我们没有使用还原公式，而是使用了 PC 和 CCPP 的乘积公式，同时还使用了 P 值和反向回归。这与逐步回归法得出的结果相比有了明显改善，而逐步回归法被认为是筛选预测因子的经典方法（Hammami 等人，2012 年）。由于大量预测因子限制了大气环流模型的数量，这种方法有助于将更多的大气环流模型和情景纳入分析，建议用于影响研究。

The main objective of the study is to analyze all the important hydrological components (i.e., precipitation, streamflow, evapotranspiration, surface flow, baseflow, snow-melt flow, TWS, and soil moisture content [SMC]) under moderate and high emission scenarios (SSP2-4.5 and SSP5-8.5) to understand their responses to changing climate in the TRHR. For the accomplishment of this study, the SDSM was used to downscale climate data of three GCMs, and the hydrological modeling system (HEC-HMS) was applied to simulate the hydrological components for the period of 1981–2100. The analyses were performed on monthly and annual basis to explore the potential impacts of climate change on the hydrological components in the near future (2021–2060) and the far future (2061–2100) relative to the baseline period (BLP) (1981–2020). On the whole, the results showed a significant increase in most hydrological components under both SSPs, which can be a good source for understanding the hydrological processes and be used for the planning and management of water resources of the highly elevated and complex region of the QTP.
本研究的主要目的是分析中度和高度排放情景（SSP2-4.5 和 SSP5-8.5）下的所有重要水文成分（即降水、溪流、蒸散、地表水流、基流、融雪水流、第三世界水系和土壤含水量 [SMC]），以了解它们对 TRHR 气候变化的响应。为完成本研究，使用 SDSM 对三个 GCM 的气候数据进行降尺度，并应用水文模拟系统（HEC-HMS）模拟 1981-2100 年期间的水文成分。分析以月度和年度为基础，探讨相对于基线期（BLP）（1981-2020 年），气候变化对近期（2021-2060 年）和远期（2061-2100 年）水文成分的潜在影响。总体而言，研究结果表明，在两个 SSP 条件下，大多数水文成分都会显著增加，这为了解水文过程提供了很好的资料，可用于规划和管理高海拔和复杂的 QTP 地区的水资源。

2 Study Area and Data Description
2 研究区域和数据说明

The TRHR, also known as the Sanjiangyuan region, is the source region of three giant rivers of Asia, that is, the transboundary Lancang (Mekong), Yangtze (Chang Jiang), and Yellow (Huang He) Rivers (Z. Wang et al., 2010). The region is located in the QTP, and its boundary stretches between 32.0–36.0°N and 89.0–103.0°E as shown in Figure 1. It covers an area of 292,700 km², where the headwaters of the Yangtze River above Zhimenda (HYaR), the Yellow River above Tangnaihai (HYeR), and the Lancang River above Xiangda (HLaR) contribute 54%, 40%, and 6% to the total area, respectively (Mahmood et al., 2020). It is an important nature reserve and a key source of fresh water in China (Jiang & Zhang, 2016). It contains a very complex mountainous geography (Jiang & Zhang, 2015), a very harsh environment, and an arid to semiarid climate (Shen et al., 2018). The TRHR is the world's largest alpine ecosystem and is composed of a high number of wetlands, rivers, lakes, glaciers, and snow-covered mountains (Tong et al., 2014). Due to the highly elevated area (i.e., 2,600–6,600 m AMSL), the water resources are highly vulnerable to climate change and global warming (Shen et al., 2018; L. Zhang et al., 2017). The region receives an annual precipitation of 262–772 mm, and an annual mean temperature varies from −5.6 to 7.8°C. The region yields 40 km³ of water, of which 22.2, 12.7, and 4.7 km³ are generated by the Yellow above Tangnaihai, Yangtze above Zhimenda, and Lancang above Xiangda, respectively (Mahmood et al., 2020).
TRHR又称三江源地区，是亚洲三大河流，即跨界澜沧江（湄公河）、长江（长江）和黄河（黄河）的源头地区（Z. Wang et al.）如图 1 所示，该区域位于青藏高原区，其边界在 32.0-36.0°N 和 89.0-103.0°E 之间。 ² 面积为 29.27 万平方公里，其中志门达以上的长江源头（HYaR）、唐乃亥以上的黄河源头（HYeR）和向达以上的澜沧江源头（HLaR）分别占总面积的 54%、40% 和 6%（Mahmood 等人，2020 年）。它是中国重要的自然保护区和重要的淡水来源（蒋和张，2016 年）。它包含非常复杂的山区地理（蒋和张，2015 年）、非常恶劣的环境和干旱到半干旱气候（沈等人，2018 年）。TRHR 是世界上最大的高山生态系统，由大量湿地、河流、湖泊、冰川和雪山组成（Tong 等人，2014 年）。由于该地区海拔较高（即 2600-6600 米 AMSL），水资源极易受到气候变化和全球变暖的影响（Shen 等人，2018 年；L. Zhang 等人，2017 年）。该地区年降水量为 262-772 毫米，年平均气温在 -5.6 至 7.8 摄氏度之间。该地区出水量为 40 千米 ³ ，其中 22.2、12.7 和 4.7 千米 ³ 分别由唐乃亥以上的黄河、止门达以上的长江和向达以上的澜沧江产生（Mahmood 等人，2020 年）。

The Hydrology and Water Resources Survey Bureau (HWRSB) of Qinghai province provided daily streamflow data for Xiangda, Zhimenda, Jimai, Maqu, and Tangnaihai hydrometric stations for the period of 1980–2015. Daily station data of precipitation (PREC), maximum temperature (Tmax), minimum temperature (Tmin), relative humidity (RH), solar radiation, and wind speed (WS) were collected from the Qinghai Meteorological Bureau (QMB), for the period of 1980–2015, shown in Figure 1 and described in Table S1 in Supporting Information S1. Since the number of climate stations is scarce in the region, the China Meteorological Forcing Data set (CMFD) was also used along with station data. Different evaluation studies such as Q. He et al. (2021) and Y. Li et al. (2022) showed better performance of CMFD as compared to other gridded data sets such as ERA5 and APHRODITE for temperature and precipitation. This is a gridded high-resolution (0.1°) near-surface meteorological data set, which is developed by combining remote sensing products, reanalysis data sets, and observed station data. It covers the whole of China and is available from 1979 to 2018 (J. He et al., 2020). The gridded data was converted into point data by taking the average of all grid points located inside each subbasin (Figure 1).
青海省水文水资源勘测局（HWRSB）提供了 1980-2015 年期间向达、致门达、吉麦、玛曲和唐乃亥水文站的日流量数据。青海省气象局收集了 1980-2015 年期间降水量（PREC）、最高气温（Tmax）、最低气温（Tmin）、相对湿度（RH）、太阳辐射和风速（WS）的日站数据，如图 1 所示，相关信息见佐证资料 S1 表 S1。由于该地区气候站数量较少，因此还使用了中国气象强迫数据集（CMFD）和气候站数据。不同的评估研究（如 Q. He 等（2021 年）和 Y. Li 等（2022 年））表明，与其他网格数据集（如 ERA5 和 APHRODITE）相比，CMFD 在温度和降水方面具有更好的性能。这是一个网格化的高分辨率（0.1°）近地面气象数据集，由遥感产品、再分析数据集和观测站数据组合而成。该数据集覆盖整个中国，时间跨度从 1979 年到 2018 年（J. He 等人，2020 年）。将网格数据转换为点数据，取每个子流域内所有网格点的平均值（图 1）。

Two kinds of large-scale data sets are required for the SDSM for the generation of future climate data: reanalysis and GCM data. The reanalysis data of 26 large-scale variables (predictors) were obtained from the National Centers for Environmental Prediction (NCEP)/National Center for Atmospheric Research (NCAR) project developed by the NCEP and the NCAR and is available for 1948–2017. All predictors are described in Table S2 in Supporting Information S1. The same predictors of three GCMs (i.e., CanESM5, MPI-ESM1.2-HR, and NorESM2-MM) were obtained from the Canadian Climate Data and Scenarios (www.climate-scenarios.canada.ca) for two Shared Socioeconomic Pathways (i.e., SSP2-4.5 and SSP5-8.5) for the historical period (1979–2014) and future period (2015–2100). The SSPs are a set of scenarios developed to explore and understand possible future socioeconomic conditions and their potential impacts on climate change and sustainable development, which were created as part of the fifth assessment report of the IPCC. SSP2-4.5 is a future scenario with a moderate emission of greenhouse gas, which is compatible with socio-economic development. Under this, effective radiative forcing is supposed to reach 4.5 W/m² by the end of the twenty-first century. SSP5-8.5 represents the future with the highest greenhouse gas emission, reaching radiative forcing to 8.5 W/m² at the end of the twenty-first century relative to pre-industrial conditions (Z. Li et al., 2022; Riahi et al., 2017). SSP2-4.5 and SSP5-8.5 represent moderately to highly vulnerable societies (Y. Zhao et al., 2022).
SDSM 生成未来气候数据需要两种大尺度数据集：再分析数据和全球大气环流模型数据。26 个大尺度变量（预测因子）的再分析数据来自美国国家环境预报中心（NCEP）/美国国家大气研究中心（NCAR）项目，该项目由美国国家环境预报中心（NCEP）和美国国家大气研究中心（NCAR）共同开发，可提供 1948-2017 年的数据。所有预测因子的描述见佐证资料 S1 表 S2。三种 GCM（即 CanESM5、MPI-ESM1.2-HR 和 NorESM2-MM）的相同预测因子来自加拿大气候数据和方案（www.climate-scenarios.canada.ca），用于历史时期（1979-2014 年）和未来时期（2015-2100 年）的两种共享社会经济路径（即 SSP2-4.5 和 SSP5-8.5）。SSP 是为探索和了解未来可能的社会经济状况及其对气候变化和可持续发展的潜在影响而制定的一套情景方案，是 IPCC 第五次评估报告的一部分。SSP2-4.5 是一种温室气体适度排放的未来情景，与社会经济发展相适应。在这种情况下，到 21 世纪末，有效辐射强迫将达到 4.5 W/m ² 。SSP5-8.5 代表温室气体排放量最高的未来，相对于工业化前的条件，21 世纪末的辐射强迫将达到 8.5 W/m ² （Z. Li 等人，2022 年；Riahi 等人，2017 年）。SSP2-4.5 和 SSP5-8.5 代表中度到高度脆弱的社会（Y. Zhao 等人，2022 年）。

The GCMs, used in this study, are the latest version of the Max Planck Institute Earth System Model (MPI-ESM), the Norwegian Earth System Model (NorESM), and the Canadian Earth System Model, which are used in CMIP6. NCEP predictors are available at 2.5° horizontal resolution while GCMs predictors are at 2.8125°. To remove the resolution mismatch, both data sets were regridded to 1.0° spatial resolution using the bilinear interpolation method, which also made an easy selection of grids for the corresponding station. Other spatial and temporal data sets, such as soil, land cover, evapotranspiration, and GRACE, used in the present study are briefly described in Table 1.
本研究中使用的 GCM 是 CMIP6 中使用的最新版马克斯-普朗克研究所地球系统模式 (MPI-ESM)、挪威地球系统模式 (NorESM) 和加拿大地球系统模式。NCEP 预测值的水平分辨率为 2.5°，而 GCMs 预测值的水平分辨率为 2.8125°。为了消除分辨率不匹配的问题，使用双线性插值法将两个数据集的空间分辨率重新划分为 1.0°，这样也便于选择相应站点的网格。表 1 简要介绍了本研究中使用的其他时空数据集，如土壤、土地覆盖、蒸散和 GRACE。

3 Methodology 3 方法

3.1 Downscaling 3.1 缩小尺度

To date, many methods and models have been developed to downscale the outputs of GCMs (Mahmood & Babel, 2013) such as bias correction, quantile mapping, automated SDSM (Hessami et al., 2008), the Long Ashton Research Station Weather Generator (Semenov & Barrow, 1997). Among them, SDSM has widely been applied throughout the world for downscaling climate variables (Fan et al., 2021; Yang et al., 2017). Many comparative studies have shown that SDSM performed relatively better than other models and/or comparable with others, as in Campozano et al. (2016), Hernanz et al. (2022), and W. Liu et al. (2013).
迄今为止，已经开发了许多方法和模型来对 GCM 的输出结果进行降尺度处理（Mahmood & Babel，2013 年），如偏差校正、量化绘图、自动 SDSM（Hessami 等人，2008 年）、Long Ashton 研究站天气生成器（Semenov & Barrow，1997 年）。其中，SDSM 已在全球范围内广泛应用于气候变量降尺度（Fan 等，2021 年；Yang 等，2017 年）。许多比较研究表明，SDSM 的性能相对优于其他模式和/或与其他模式相当，如 Campozano 等（2016 年）、Hernanz 等（2022 年）和 W. Liu 等（2013 年）。

SDSM developed by Wilby et al. (2002) is a combination of MLR and the stochastic weather generator (SWG). MLR is applied to establish a statistical relationship between local scale (gauge data) climate variables (e.g., PREC, Tmax, Tmin, WS, and RH) and large scale variables such PREC, sea level pressure, and meridional windspeed of NCEP during the calibration process. SWG uses the calibrated parameters and simulates climate data for the historical as well as future using NCEP and GCM predictors (Mahmood & Babel, 2013; Wilby et al., 2014). Three data sets are required to configure this model: station data, NCEP, and GCM predictors (Wilby et al., 2014). Station data and NCEP predictors are used to calibrate the model, and GCM predictors are forced into the calibrated SDSM to simulate future climate data. There are three main steps in this method: screening predictors, calibration and validation, and data simulation for the future.
Wilby 等人（2002 年）开发的 SDSM 是 MLR 和随机天气生成器（SWG）的结合。在校准过程中，应用 MLR 在局部尺度（测量数据）气候变量（如 PREC、Tmax、Tmin、WS 和 RH）和大尺度变量（如 PREC、海平面气压和 NCEP 的经向风速）之间建立统计关系。SWG 使用校准参数，并利用 NCEP 和 GCM 预测器模拟历史和未来的气候数据（Mahmood & Babel，2013 年；Wilby 等人，2014 年）。配置该模型需要三个数据集：站点数据、NCEP 和 GCM 预测因子（Wilby 等人，2014 年）。站点数据和 NCEP 预测因子用于校准模型，而 GCM 预测因子则被强制输入校准后的 SDSM，以模拟未来的气候数据。该方法有三个主要步骤：筛选预测因子、校准和验证以及未来数据模拟。

The screening is the central and most important process for all statistical downscaling techniques. In SDSM, the predictors are selected based on the correlation matrix, P-value, and explained variance. However, the screening process in SDSM is very tedious, time-consuming, subjective, and also requires expertise to select some suitable predictors (Hammami et al., 2012; Mahmood & Babel, 2013). For example, a large number of predictors in regression can cause multi-collinearity and a small number may not be enough to explain the local variable, resulting in poor results. Mandal et al. (2016) explored that the statistical downscaling methods often suffer from the multi-collinearity of the predictors, especially, in the case of MLR. The ordinary least square estimations of regression coefficients may be unstable due to the collinearity among predictors (P. Singh et al., 2023). To overcome this issue, we used the following method:
筛选是所有统计降尺度技术的核心和最重要的过程。在 SDSM 中，预测因子的选择基于相关矩阵、P 值和解释方差。然而，SDSM 的筛选过程非常繁琐、耗时、主观，还需要专业知识来选择一些合适的预测因子（Hammami 等人，2012 年；Mahmood & Babel，2013 年）。例如，回归中大量的预测因子会导致多重共线性，而少量的预测因子可能不足以解释局部变量，从而导致结果不佳。Mandal 等人（2016 年）研究发现，统计降尺度方法经常会受到预测因子多重共线性的影响，尤其是在 MLR 的情况下。由于预测因子之间的共线性，回归系数的普通最小二乘法估计可能不稳定（P. Singh 等人，2023 年）。为了克服这一问题，我们采用了以下方法：

3.1.1 Proposed Screening Method
3.1.1 建议的筛选方法

In this study, we used a combination of correlation, PC, p-value, and backward regression, which will automatically deal with collinearity, to provide an optimum set of predictors, and rank them accordingly. This is the modified form of the approach used by Mahmood and Babel (2013). The major steps of the proposed method are outlined in Figure 2, and are described below:
在本研究中，我们结合使用了相关性、PC、P 值和反向回归，这将自动处理共线性问题，提供一组最佳预测因子，并对它们进行相应的排序。这是 Mahmood 和 Babel（2013 年）所用方法的改进形式。拟议方法的主要步骤如图 2 所示，下文将对其进行说明：

1. Make a correlation matrix between predictand and predictors. Highlight the predictors having high correlations (e.g., >0.78–0.85) and remove the predictors having a low correlation with predictand. For example, the correlation between predictor X and predictor Y is 0.98, and they have correlations of 0.70 and 0.60, respectively, with predictand. So, the predictor Y will be removed for further analysis. Similarly, do it for others. In this study, we used 0.78–0.85 as critical correlation (i.e., 0.6–0.7 R-squared) because different studies have suggested using the variance inflation factor less than 2.5–3.5 to reduce the collinearity effect (Johnston et al., 2018).
   制作预测因子和预测因子之间的相关矩阵。突出显示相关性较高的预测因子（如大于 0.78-0.85），删除与预测因子相关性较低的预测因子。例如，预测因子 X 和预测因子 Y 的相关性为 0.98，而它们与预测对象的相关性分别为 0.70 和 0.60。因此，进一步分析时将删除预测因子 Y。同样，对其他预测因子也是如此。在本研究中，我们使用 0.78-0.85 作为临界相关性（即 0.6-0.7 R 平方），因为不同的研究建议使用小于 2.5-3.5 的方差膨胀因子来减少共线性效应（Johnston et al.）

2. After removing highly correlated predictors, we removed statistically insignificant predictors. For this, we used the backward regression analysis using the P-value criterion, at a significance level of 0.05. Other criteria such as the Akaike information criterion, Bayesian information criterion, or R-square can also be used. In this procedure, the null hypothesis was no relationship (no correlation) between predictor and predictand and the alternative hypothesis was a strong relationship (strong correlation) between predictand and predictor. This means that any P-value greater than 0.05 between predictor and predictand shows that there is no effect of a predictor on predictand and any P-value less than 0.05 shows a strong relationship between predictor and predictand. The stepwise backward PC process can also be used to eliminate the insignificant predictors.
   在剔除高度相关的预测因子后，我们又剔除了在统计上不重要的预测因子。为此，我们使用 P 值标准进行反向回归分析，显著性水平为 0.05。也可以使用其他标准，如 Akaike 信息标准、贝叶斯信息标准或 R 方差。在这一过程中，零假设是预测因子和预测对象之间没有关系（无相关性），而备择假设是预测对象和预测因子之间有很强的关系（强相关性）。这意味着，预测因子和预测对象之间的任何 P 值大于 0.05，都表明预测因子对预测对象没有影响，而任何 P 值小于 0.05，都表明预测因子和预测对象之间有很强的关系。也可以使用逐步后向 PC 流程来剔除不显著的预测因子。

3. In this step, we used a product of correlation and PC along with P-value (at α = 0.05) to choose optimal predictors and to rank them. For this, first, all the predictors (selected from the previous steps) were arranged in descending order based on the correlations between predictors and predictand (RPP). The predictor at the top was referred to as a supper predictor (SP). Second, we determined the PC of each of the predictors one by one, taking away the effect of SP, along with P-values. Third, we calculated the product of the correlation coefficient and PC (PCP). Finally, the predictor having the highest PCP value was selected as second SP, and predictors having a P-value greater than 0.05 were removed. To get the third SP, the PCs were calculated by taking away the effect of the first SP and second SP. This process was repeated until all predictors were ranked according to their effectiveness. The p-value was used to remove insignificant predictors in each repetition. The product formula along with P-value can be a very effective and simple tool to rank the most effective predictors and remove insignificant predictors.
   在这一步中，我们使用相关性和 PC 的乘积以及 P 值（α = 0.05）来选择最佳预测因子并对其进行排序。为此，首先根据预测因子与预测对象之间的相关性（RPP），将所有预测因子（从之前的步骤中选出）按降序排列。排在最前面的预测因子被称为辅助预测因子（SP）。其次，在剔除 SP 的影响后，我们逐一确定每个预测因子的 PC 值以及 P 值。第三，计算相关系数与 PC 的乘积（PCP）。最后，选择 PCP 值最高的预测因子作为第二个 SP，并剔除 P 值大于 0.05 的预测因子。为了得到第三个 SP，在计算 PC 时要剔除第一个 SP 和第二个 SP 的影响。此过程重复进行，直到所有预测因子都按其有效性排序。每次重复计算时，都会使用 P 值来剔除不显著的预测因子。乘积公式和 P 值是一种非常有效和简单的工具，可用于对最有效的预测因子进行排序，并剔除不重要的预测因子。

4. This method was compared with the stepwise regression for evaluation, which is considered a classical method for the screening of predictors (Hammami et al., 2012).
   该方法与逐步回归评估法进行了比较，后者被认为是筛选预测因子的经典方法（Hammami 等人，2012 年）。

3.1.2 Calibration and Validation
3.1.2 校准与验证

Based on the available daily station data, SDSM was calibrated for 1980–2005 and validated for 2006–2015 for each climate variable (i.e., PREC, Tmax, Tmin, WS, and RH) for all stations and CMFD in the TRHR. There are two modules in SDSM for the calibration of the model, that is, annual and monthly. The annual model develops a single regression model for the whole time series while the monthly module generates 12 regression models, one for each month separately (Huang et al., 2011). In this study, the monthly module was used during the calibration process because according to Mahmood and Babel (2013), the monthly module provides better results than the annual module. In SDSM, there are two modules (i.e., conditional and unconditional), which are selected according to the variable. The unconditional module is used for independent or unconditional variables such as temperature and windspeed and the conditional for conditional variables such as PREC (Wilby et al., 2002). Since PREC data often does not follow normality (normal distribution), it requires transformation before using it in the calibration process to get good results (Khan et al., 2006). There are many transformation methods in SDSM such as lambda, log, ln, X², X^1/4, and X^1/2. In the present study, we evaluated all transformations and found fourth root and lambda providing high values of coefficient of determination (Table S3 in Supporting Information S1), fourth root transformation has also been used in Huang et al. (2011) and Khan et al. (2006). Figure 3 shows monthly PREC simulated at Wudaoliang station using no-transformation, lambda, and fourth root. With no transformation, the simulated PREC was well overestimated in all months. Nonetheless, the simulated PREC with transformations captured monthly variations better than no transformation. The coefficient of determination (R²) and root mean square error (RMSE) were applied for the evaluation of SDSM.
根据现有的每日站点数据，SDSM 对 TRHR 所有站点和 CMFD 的每个气候变量（即 PREC、Tmax、Tmin、WS 和 RH）进行了 1980-2005 年的校核和 2006-2015 年的验证。SDSM 有两个模块用于校准模式，即年度和月度模块。年度模块为整个时间序列建立一个回归模型，而月度模块则为每个月分别建立 12 个回归模型（Huang 等，2011 年）。本研究在校准过程中使用了月度模块，因为根据 Mahmood 和 Babel（2013 年）的研究，月度模块比年度模块提供了更好的结果。在 SDSM 中，有两个模块（即有条件模块和无条件模块），可根据变量选择。无条件模块用于温度和风速等独立或无条件变量，有条件模块用于 PREC 等有条件变量（Wilby 等，2002 年）。由于 PREC 数据通常不符合正态性（正态分布），因此在校准过程中使用之前需要对其进行转换，以获得良好的结果（Khan 等人，2006 年）。SDSM 中有许多转换方法，如 lambda、log、ln、X ² 、X ^1/4 和 X ^1/2 等。在本研究中，我们对所有变换进行了评估，发现第四根和 lambda 提供了较高的决定系数值（佐证资料 S1 表 S3），黄等人（2011 年）和 Khan 等人（2006 年）也使用了第四根变换。图 3 显示了采用无变换、λ 和第四根变换模拟的五道梁站月度 PREC。在不进行变换的情况下，各月模拟的 PREC 都被高估了。 尽管如此，与不进行转换相比，经过转换的模拟 PREC 能更好地捕捉月度变化。在评估 SDSM 时采用了判定系数（R ² ）和均方根误差（RMSE）。

3.2 HEC-HMS's Setup 3.2 HEC-HMS 的设置

The Hydrological Modeling System of Hydrologic Engineering Center (HEC-HMS) (William & Fleming, 2010) is a semi-distributed hydrological model and has been used throughout the world for different hydrologic applications such as flood modeling, water resource assessment, climate change impacts assessment, urban flooding, flood warning system, stream restoration, water availability, and streamflow forecasting, as in Mahmood and Jia (2017), Ramly and Tahir (2016), and Zema et al. (2017). There are four main components for the setup of HEC-HMS: the meteorological model (MM), basin model (BM), control specification (CS), and data manager (DM) (Verma et al., 2010). The BM stores the physical characteristics (e.g., areas, lengths, and slopes) of a catchment, which can be extracted from Digital Elevation Model (DEM). The MM calculates the spatial distribution of climate variables over a catchment; the CS is used to specify a simulation period to run the model; and the DM stores and manages time-series data (e.g., temperature, WS, and PREC) (William & Fleming, 2010).
水文工程中心水文建模系统（HEC-HMS）（William & Fleming，2010 年）是一个半分布式水文模型，已在世界各地用于不同的水文应用，如洪水建模、水资源评估、气候变化影响评估、城市洪水、洪水预警系统、溪流恢复、水资源可用性和溪流预报，如 Mahmood 和 Jia（2017 年）、Ramly 和 Tahir（2016 年）以及 Zema 等（2017 年）。HEC-HMS 的设置有四个主要组成部分：气象模型（MM）、流域模型（BM）、控制规范（CS）和数据管理器（DM）（Verma 等人，2010 年）。流域模型存储流域的物理特征（如面积、长度和坡度），这些特征可从数字高程模型（DEM）中提取。MM 计算集水区气候变量的空间分布；CS 用于指定运行模型的模拟期；DM 存储并管理时间序列数据（如温度、WS 和 PREC）（William & Fleming，2010 年）。

For the setup of the model, we included Thiessen polygon, Penman-Monteith, and Temperature Index Module (TIM) in MM and soil moisture accounting (SMA), Clark Unit Hydrograph, Muskingum Channel Routing, Linear Reservoir Baseflow, Dynamic Canopy, and Surface Storage methods in the BM, as shown in Figure 5. SMA is an advanced continuous and complex method in HEC-HMS to calculate losses and excess rainfall (Bhuiyan et al., 2017). It simulates the movement of water over time through a set of storage zones in the groundwater and soil profile layers (William & Fleming, 2010). The SMA-algorithm represents the watershed with five layers (i.e., canopy storage, surface storage, soil profile storage, groundwater layer 1, and groundwater layer 2), and it requires 13 parameters (e.g., surface depression storage, canopy interception storage, soil storage, and infiltration rate), which are estimated during the calibration process. Thiessen polygon is used to interpolate climate data over the watershed by providing weights to climate stations according to the covering area. The penman-Monteith method is used to measure the evapotranspiration in the basin. TIM estimates the streamflow from snowfall for each subbasin. In the present study, each subbasin was divided into 3–5 elevation bands to increase the accuracy of simulating streamflow from snow-melt because the region has a complex topography with highly varying elevations. The other methods are comprehensively described in Feldman (2000).
在模型设置中，我们在 MM 中加入了 Thiessen 多边形、Penman-Monteith 和温度指数模块 (TIM)，在 BM 中加入了土壤水分核算 (SMA)、克拉克单元水文图、马斯金姆河道路径、线性水库基流、动态冠层和地表蓄水方法，如图 5 所示。SMA 是 HEC-HMS 中的一种高级连续复杂方法，用于计算损失和过量降雨（Bhuiyan 等人，2017 年）。它通过地下水和土壤剖面层中的一组储水区模拟水流随时间的运动（William & Fleming，2010 年）。SMA 算法用五个层（即冠层存储、地表存储、土壤剖面存储、地下水层 1 和地下水层 2）表示流域，需要 13 个参数（如地表凹陷存储、冠层截流存储、土壤存储和渗透率），这些参数在校准过程中进行估算。Thiessen 多边形用于对流域内的气候数据进行插值，根据覆盖面积对气候站进行加权。Penman-Monteith 法用于测量流域的蒸散量。TIM 根据降雪量估算每个子流域的流量。在本研究中，由于该地区地形复杂，海拔高度变化很大，因此将每个子流域划分为 3-5 个海拔带，以提高模拟融雪产生的溪流的精度。其他方法在 Feldman（2000 年）中有全面介绍。

There are two kinds of model parameters: physical and process parameters (Kan et al., 2019; Q. Zhang et al., 2008). Physical parameters represent the physical features (e.g., basin area and river length) of a catchment (Q. Zhang et al., 2008), which were obtained by delineating the SRTM-DEM in this study. There was a total of 22 process parameters for all methods, as shown in Figure 5. Generally, these parameters cannot be measured directly from a catchment, and these are estimated indirectly through model calibration (Q. Zhang et al., 2008). The initial values of these parameters are needed to run the model (Kan et al., 2019). However, these initial values must be rational and logical, otherwise, these can cause misleading results. Therefore, a systematic and comprehensive approach, as in Mahmood and Jia (2022), was applied to estimate these values from soil and land cover data sets. This not only reduced the time to calibrate the model but also helped in the acquisition of realistic parameters during calibration.
模型参数分为两种：物理参数和过程参数（Kan 等人，2019 年；Q. Zhang 等人，2008 年）。物理参数代表集水区的物理特征（如流域面积和河流长度）（Q. Zhang et al.如图 5 所示，所有方法共有 22 个过程参数。一般来说，这些参数无法从流域中直接测量，需要通过模型校准来间接估算（Q. Zhang 等，2008 年）。运行模型需要这些参数的初始值（Kan 等，2019 年）。然而，这些初始值必须合理、合乎逻辑，否则会造成误导性结果。因此，与 Mahmood 和 Jia（2022 年）一样，我们采用了一种系统而全面的方法，从土壤和土地覆被数据集中估算这些值。这不仅缩短了校准模型的时间，还有助于在校准过程中获取真实的参数。

3.2.1 Calibration, Validation, and Sensitivity Analysis
3.2.1 校准、验证和敏感性分析

In the present study, the split sample approach was applied for the calibration of HEC-HMS, which is a classic method and central to a hierarchical scheme for validating hydrological models. For this, the whole period was divided into three parts 1981–2005, 2006–2010, and 2011–2015. The middle one (2006–2010) was used for calibration because of the least missing observations, and the other two were used for validation to make a robust evaluation of the model. The model was calibrated and validated at five available gauges: Xiangda at the Lancang River, Zhimenda at the Yangtze River, Jimai, Maqu, and Tangnaihai at the Yellow River.
在本研究中，对 HEC-HMS 的校核采用了分割样本方法，这是一种经典方法，也是水文模型分层验证方案的核心。为此，整个时期被分为 1981-2005 年、2006-2010 年和 2011-2015 年三个部分。由于中间部分（2006-2010 年）缺失的观测资料最少，因此将其用于校核，而其他两个部分则用于验证，以便对模型进行稳健的评估。该模型在五个现有测站进行了校核和验证：模型在五个现有测站进行了校核和验证：澜沧江的象达测站、长江的支门达测站、黄河的吉迈测站、玛曲测站和唐乃亥测站。

Before calibration, we performed the sensitivity analysis, as in Belayneh et al. (2020) and Mahmood and Jia (2019). Sensitivity analysis explores the important, sensitive, and influential parameters in a watershed and triggers the calibration process, especially in the case of manual calibration (Devak & Dhanya, 2017; Mahmood & Jia, 2022). In the present study, the sensitivity analyses were carried out by changing each parameter by 10% each time (between −40% and +40%) to observe the changes in simulated discharge. Two indicators, Peak Flow (PF) and Total Flow Volume (TFV) were used to observe the effect of each parameter on streamflow, which are mostly used in literature, as in Acheampong et al. (2023), Mahmood et al. (2020), and Palacios-Cabrera et al. (2022). In the case of PF, the top four sensitive parameters were maximum infiltration, storage coefficient, soil percolation, and surface storage; and in the case of TFV, these were the percolations and storage coefficients of groundwater layer 1 and groundwater layer 2, as shown in Figure 4.
在校准之前，我们进行了敏感性分析，如 Belayneh 等人（2020 年）和 Mahmood 与 Jia（2019 年）。灵敏度分析可探索流域中重要、敏感和有影响的参数，并触发校准过程，尤其是在手动校准的情况下（Devak 和 Dhanya，2017 年；Mahmood 和 Jia，2022 年）。在本研究中，通过每次改变每个参数 10%（-40% 到 +40%）来进行敏感性分析，以观察模拟排水量的变化。使用峰值流量（PF）和总流量（TFV）这两个指标来观察各参数对溪流的影响，文献中大多使用这两个指标，如 Acheampong 等人（2023 年）、Mahmood 等人（2020 年）和 Palacios-Cabrera 等人（2022 年）。对于 PF，前四个敏感参数分别是最大入渗量、储水系数、土壤渗流和地表储水；对于 TFV，前四个敏感参数分别是地下水层 1 和地下水层 2 的渗流和储水系数，如图 4 所示。

In addition to the model evaluation with streamflow data, the model was also evaluated for other hydrological components (e.g., baseflow, SMC, and TWS), because our objective was to study all important hydrological components in the region. However, the observed data of these components were not available in the region. Therefore, we obtained reanalysis and remote sensing data for SMC, AET, SWE, and TWS from freely available sources, described in Table 1.
除了利用河水流量数据对模型进行评估外，还对模型的其他水文成分（如基流、SMC 和 TWS）进行了评估，因为我们的目标是研究该地区所有重要的水文成分。但是，该地区没有这些成分的观测数据。因此，我们从表 1 所述的免费来源获取了 SMC、AET、SWE 和 TWS 的再分析和遥感数据。

3.2.2 Estimating Baseflow, TWS Changes, and AET
3.2.2 估算基流、TWS 变化和 AET

Baseflow is one of the most important parameters of a hydrological cycle, however, difficult to measure in the watershed (Miller et al., 2014), especially located in harsh climates like the TRHR, where it is even very difficult to manage the streamflow gauges. Mostly, baseflow separation techniques are used to separate the baseflow from streamflow, as described in Brodie and Hostetler (2005), Murphy et al. (2009), and Nathan and McMahon (1990). A frequently used digital filter, the recursive digital filter developed by Nathan and McMahon (1990), was applied in this study, as below:
基流是水文循环中最重要的参数之一，但在流域中却很难测量（米勒等人，2014 年），尤其是像 TRHR 这样气候恶劣的地区，甚至很难管理流量计。大多数情况下，基流分离技术被用来将基流从溪流中分离出来，如 Brodie 和 Hostetler（2005 年）、Murphy 等人（2009 年）以及 Nathan 和 McMahon（1990 年）所述。本研究采用了一种常用的数字滤波器，即 Nathan 和 McMahon（1990 年）开发的递归数字滤波器，具体如下：

$$${f}_{k}=\alpha {f}_{k-1}+\frac{(1+\alpha )}{2}\left[{y}_{k}-{y}_{k-1}\right]$$$ (1)

where f_k is quick flow at the kth time, α is a filter parameter, and y_k is streamflow. An α value of 0.925 for daily flow and 0.995 for hourly flow is recommended by Nathan and McMahon (1990).
其中，f _k 是第 k 个时间点的快速流量，α 是滤波参数，y _k 是溪流流量。Nathan 和 McMahon（1990 年）建议日流量的 α 值为 0.925，小时流量的 α 值为 0.995。

TWS is crucial for global as well as regional hydrological cycles and water resources management (Chen et al., 2017). It includes canopies, snow/ice, rivers, lakes, wetlands, soil, and groundwater and is a critical component of the water and energy budget (Pokhrel et al., 2021). However, it is difficult to measure TWS directly (Xu, 2017). Xiong et al. (2022) estimated TWS by summing the SM and SWE simulated by GCMs. In this study, TWS was estimated by summing the canopy storage (S_canopy), surface storage (S_surface), soil storage (S_soil), groundwater storage from layer 1 and layer 2 (S_ground1, S_ground2), and water stored in snow as SWE (S_swe); and AET, another one of the most important hydrological components, was calculated by the water balance equation, as given below:
TWS 对于全球以及区域水文循环和水资源管理至关重要（Chen 等人，2017 年）。它包括树冠、冰雪、河流、湖泊、湿地、土壤和地下水，是水和能量预算的重要组成部分（Pokhrel 等，2021 年）。然而，很难直接测量 TWS（Xu，2017 年）。Xiong 等人（2022 年）通过将 GCM 模拟的 SM 和 SWE 相加来估算 TWS。在本研究中，TWS 是通过将冠层蓄水量（S _canopy ）、地表蓄水量（S _surface ）、土壤蓄水量（S _soil ）、第 1 层和第 2 层地下水蓄水量（S _ground1 、S _ground2 ）以及作为 SWE 的雪中蓄水量（S _swe ）相加而估算的；而另一个最重要的水文成分 AET 是通过水平衡方程计算的，如下所示：

$$$\text{TWS}={S}_{\text{canopy}}+{S}_{\text{surface}}+{S}_{\text{soil}}+{S}_{\text{ground}1}+{S}_{\text{ground}2}+{S}_{\text{swe}}$$$ (2)

$$$\text{AET}=P-{Q}_{\text{sim}}-{\increment}S$$$ (3)

where P, Q_sim, and ∆S represent PREC, simulated flow, and changes in water storage, respectively, in the basin. Mostly ∆S is assumed negligible for analysis over a longer period (≥10 years) (Zhu et al., 2019) because it is difficult to measure. Nonetheless, in the present study, ∆S was estimated from TWS, as below:
其中，P、Q _sim 和 ∆S 分别表示流域的 PREC、模拟流量和蓄水量变化。由于 ∆S 难以测量，在较长时期（≥10 年）的分析中，通常假定其可以忽略不计（Zhu 等，2019 年）。然而，在本研究中，∆S 是通过 TWS 估算的，如下所示：

$$${\increment}S={\text{TWS}}_{t}-{\text{TWS}}_{t-1}$$$ (4)

Five statistical indicators (described below), Nash-Sutcliffe efficiency (E), coefficient of determination (R²), RMSE, normalized root mean square error (NRMSE), percent bias (PBIAS), and graphs were used for the model evaluation. The whole procedure (e.g., input, outputs, hydrologic parameters, calibration, and validation) for the setup of HEC-HMS for the TRHR is presented in Figure 5. After successful calibration and validation of SDSM and HEC-HMS, climate data (i.e., PREC, Tmax, Tmin, RH, and WS) were simulated by SDSM for 1981–2100, which was then used in HEC-HMS to simulate streamflow, surface flow, baseflow, snowmelt water, SMC, AET, and TWS for the period of 1981–2100. The whole data was divided into three equal periods for further analysis: 1981–2020, 2021–2060, and 2061–2100. The period 1981–2020 was used as the BLP to assess the relative changes in hydrological components for the near future period (NFP, 2021–2060) and far future period (FFP, 2061–2100).
模型评价采用了五项统计指标（如下所述）：纳什-苏特克利夫效率（E）、判定系数 （R ² ）、均方根误差（RMSE）、归一化均方根误差（NRMSE）、偏差百分比 （PBIAS）和图形。图 5 显示了为 TRHR 设置 HEC-HMS 的整个过程（如输入、输出、水文参数、校核和验证）。在 SDSM 和 HEC-HMS 成功校核和验证后，SDSM 模拟了 1981-2100 年的气候数据（即 PREC、Tmax、Tmin、RH 和 WS），然后 HEC-HMS 利用这些数据模拟了 1981-2100 年期间的河水流量、地表水流量、基流、融雪水量、SMC、AET 和 TWS。为进一步分析，将整个数据分为三个相等的时段：1981-2020 年、2021-2060 年和 2061-2100 年。以 1981-2020 年为基准期，评估近期（NFP，2021-2060 年）和远期（FFP，2061-2100 年）水文成分的相对变化。

$$$E=1-\frac{\sum {\left({Q}_{\text{sim}}-{Q}_{\text{obs}}\right)}^{2}}{\sum {\left({Q}_{\text{obs}}-\overline{{Q}_{\text{obs}}}\right)}^{2}}$$$

$$${R}^{2}=\frac{\sum \left({Q}_{\text{obs}}-\overline{{Q}_{\text{obs}}}\right)\times \left({Q}_{\text{sim}}-\overline{{Q}_{\text{sim}}}\right)}{\sqrt{\sum {\left({Q}_{\text{obs}}-\overline{{Q}_{\text{obs}}}\right)}^{2}\times {\left({Q}_{\text{sim}}-\overline{{Q}_{\text{sim}}}\right)}^{2}}}$$$

$$$\text{PBIAS}(\%)=100\times \sum {Q}_{\text{sim}}-{Q}_{\text{obs}}/\sum {Q}_{\text{obs}}$$$

$$$\text{NRMSE}=\frac{\text{RMSE}}{{\sigma }_{\text{obs}}}=\frac{\sqrt{\left(\frac{1}{n}\sum\limits _{i=1}^{n}{\left({Q}_{\text{sim}}-{Q}_{\text{obs}}\right)}^{2}\right)}}{{\sigma }_{\text{obs}}}$$$

where Q_obs, Q_sim, and σ describe observed streamflow, simulated streamflow, and standard deviation (SD) of observed data, respectively.
其中 Q _obs 、Q _sim 和 σ 分别描述观测到的流量、模拟流量和观测数据的标准偏差 (SD)。

4 Results 4 成果

4.1 SDSM's Performance Evaluation
4.1 战略部署物资管理司的绩效评估

4.1.1 Evaluation of Screening Method
4.1.1 筛选方法评估

The proposed screening method (PSM) was evaluated with the stepwise regression screening method (SRSM), which is considered a classical method. Table 2 shows the predictors selected by the SRSM and PSM for Tmax and PREC at different sites in the TRHR. The SRSM selected 12–18 predictors for each site for Tmax and PREC while the PSM only 4–6 predictors for each variable, removing insignificant predictors, which minimized the effect of multi-collinearity in the regression model. To check the performance of selected predictors, 16 predictors selected by SRSM and four by the PSM (Table 2) out of 26 NCEP predictors for PREC at Ruoergai were used in SDMS to simulate PREC, and the simulations were compared with observations, as shown in Figure 6. This shows that even the number of predictors selected by PSM were much smaller, the simulation results were well comparable with the SRSM. Nonetheless, the correlation coefficient was a little higher in the case of SRSM (i.e., 0.89 by SRSM and 0.87 by this method), which might be due to multi-collinearity. The multi-collinearity produces the variance of the coefficient estimates, and the increased variance makes the coefficient estimates very sensitive to monitor changes in the data, which produces unstable coefficients (P. Singh et al., 2023).
拟议的筛选方法（PSM）与逐步回归筛选方法（SRSM）进行了评估，后者被认为是一种经典方法。表 2 显示了 SRSM 和 PSM 为 TRHR 不同地点的 Tmax 和 PREC 选择的预测因子。SRSM 为每个地点的 Tmax 和 PREC 选择了 12-18 个预测因子，而 PSM 则为每个变量只选择了 4-6 个预测因子，剔除了不重要的预测因子，从而将回归模型中多重共线性的影响降至最低。为了检验所选预测因子的性能，SDMS 利用 SRSM 和 PSM 分别从 NCEP 针对若尔盖 PREC 的 26 个预测因子中选出的 16 个和 4 个预测因子（表 2）对 PREC 进行了模拟，并将模拟结果与观测结果进行了比较，如图 6 所示。从图 6 可以看出，即使 PSM 选用的预报因子数量更少，模拟结果也与 SRSM 相当。不过，SRSM 的相关系数略高（即 SRSM 为 0.89，本方法为 0.87），这可能是由于多重共线性造成的。多重共线性导致系数估计值的方差增大，方差增大使得系数估计值对数据监测变化非常敏感，从而导致系数不稳定（P. Singh 等人，2023 年）。

Table 2. Selected National Centers for Environmental Prediction Predictors for Maximum Temperature and Precipitation at Different Sites in the Three-River Headwater Region, Using Stepwise Regression and the Proposed Method
表 2.美国国家环境预测中心采用逐步回归法和拟议方法对三河源头地区不同地点最高气温和降水量的选定预测值

	Wudaoliang 五道梁	Tuotuo	Qumalai 库马莱	Zhiduo 智多	Zaduo 扎朵	Nangqian 南迁	Maduo 马多	Ruoergai 若尔盖	Maqu
Selected predictors by stepwise regression 逐步回归法选定的预测因子
For maximum temperature 最高温度
1	dswr	dswr	dswr	dswr	dswr	dswr	dswr	dswr	dswr
2	lftx	lftx	lftx	lftx	lftx	lftx	lftx	lftx	lftx
3	mslp	mslp	mslp	p5_u	mslp	p5_u	mslp	mslp	mslp
4	p5_u	p5_f	p5_f	p5_z	p5_u	p5_z	p5_u	p500	p5_u
5	p5_z	p5_u	p5_u	p5th 第5页	p5_v	p5th 第5页	p5_z	p5_u	p8_f
6	p5th 第5页	p5_z	p5_z	p5zh	p5_z	p8_f	p5zh	p5_z	p__f
7	p5zh	p8_f	p8_f	p8_v	p8_f	p8_v	p8_z	p5zh	p__u
8	p8_f	p8_z	p8_z	p8_z	p8_z	p8_z	p8th 第8页	p8_u	p__z
9	p8_z	p8zh	p8zh	p__f	p8th 第8页	p8th 第8页	p8zh	p8th 第8页	p_zh
10	p8th 第8页	p__f	p__f	p__v	p__f	p__f	p__f	p__f	prec
11	p__f	p__u	p__u	p__z	p__v	p__z	p__u	p__z	r500
12	p__u	p__v	p__v	p_th	p_th	p_zh	p__v	p_th	r850
13	p__z	p__z	p__z	prec 前兆	p_zh	pr_wtr	p__z	p_zh	temp
14	p_zh	p_th	p_th	r500	prec 前兆	prec 前兆	p_th	prec 前兆
15	prec 前兆	r500	r500	rhum 朗姆酒	rhum 朗姆酒	rhum 朗姆酒	prec 前兆	r850
16	rhum 朗姆酒	r850	r850	temp 温度	temp 温度	temp 温度	r500	temp 温度
17	temp 温度	rhum 朗姆酒	rhum 朗姆酒				rhum 朗姆酒
18		temp 温度	temp 温度				temp 温度
For precipitation 降水量
1	dswr	dswr	dswr	dswr	lftx	lftx	dswr	dswr	dswr
2	lftx	lftx	lftx	lftx	mslp	mslp	lftx	lftx	lftx
3	mslp	mslp	mslp	mslp	p500	p500	mslp	mslp	p500
4	p500	p500	p5_z	p5_z	p5_f	p5_f	p5_f	p500	p5_z
5	p5_u	p5_z	p5zh	p5zh	p5_z	p5_z	p5_z	p5_f	p5zh
6	p5_z	p5th 第5页	p8_u	p8_u	p5th 第5页	p5th 第5页	p5zh	p5_z	p8_f
7	p5th 第5页	p5zh	p8_z	p8_z	p5zh	p5zh	p8_f	p5th 第5页	p8_u
8	p5zh	p8_u	p8th 第8页	p8th 第8页	p8_f	p8_v	p8_v	p5zh	p8_z
9	p8_u	p8_z	p__f	p__f	p8_z	p8_z	p8_z	p8_u	p8th
10	p8_z	p__v	p__z	p__z	p8th 第8页	p8th 第8页	p8th 第8页	p8th 第8页	p__f
11	p8th 第8页	p__z	p_zh	pr_wtr	p__f	p__f	p__f	p__f	pr_wtr
12	p__f	p_zh	pr_wtr	prec 前兆	p__v	p__v	p__u	p__z	prec
13	p__v	pr_wtr	prec 前兆		p__z	p__z	p__v	pr_wtr	r500
14	p__z	prec 前兆			pr_wtr	pr_wtr	pr_wtr	prec 前兆
15	prec 前兆				prec 前兆	prec 前兆	prec 前兆	r500
16	pr_wtr					r850	r500	r850
17	shum 避						r850
Selected predictors by the proposed method 通过拟议方法选出的预测因子
For maximum temperature 最高温度
1	temp 温度	temp 温度	temp 温度	temp 温度	temp 温度	temp 温度	temp 温度	temp 温度	temp
2	lftx	lftx	lftx	lftx	lftx	lftx	lftx	lftx	lftx
3	dswr	mslp	mslp	mslp	dswr	dswr	dswr	mslp	p5_u
4	p_zh	r500	p5_u	p8_v	p5_u	p5_u	p__f	p500	r500
5	p__z	p5_u	p5_f	p5_u	rhum 朗姆酒		mslp	p5_u	p_zh
6	p8_z	p5_f	p__z	p__z			p5_u		mslp
For precipitation 降水量
1	prec 前兆	pr_wtr	pr_wtr	pr_wtr	pr_wtr	pr_wtr	pr_wtr	pr_wtr	pr_wtr
2	p500	p8_z	p8_u	p8_u	prec 前兆	prec 前兆	prec 前兆	prec 前兆	prec
3	lftx	p5zh	p8_z	p5zh	lftx	r850	lftx	p8_u	r500
4	mslp	p__v	p5zh	p8_z	p500	p__v	p5_f	p__u	lftx
5	p__v	p500		mslp	p5_f	lftx	r850		p500
6	p5_u				p5_z

4.1.2 Evaluation of SDSM and GCMs
4.1.2 SDSM 和 GCM 的评估

After the successful screening of NCEP predictors using the PSM, SDSM was calibrated for PREC, Tmax, Tmin, RH, and WS at more than 100 points (Figure 1) for 1981–2005 and validated for 2006–2015. Table 3 shows the maximum, minimum, and mean values of correlation coefficients (R), coefficient of determination (R²), and RMSE calculated at all sites in the TRHR, using monthly time series. On average, R and R² values were higher than 0.89 and 0.80 for all variables except WS, which had average values of 0.6 and 0.67, respectively. The values of RMSE for all the variables were also well within the acceptable range (Table 3). SDSM showed the best performance for temperature and relatively bad for WS. It was also observed that validation results were even a little better than the calibration.
在利用 PSM 成功筛选出 NCEP 预测因子后，SDSM 在 100 多个点（图 1）对 1981-2005 年的 PREC、Tmax、Tmin、RH 和 WS 进行了校准，并对 2006-2015 年进行了验证。表 3 列出了 TRHR 所有站点利用月时间序列计算的相关系数（R）、判定系数（R ² ）和均方根误差的最大值、最小值和平均值。除 WS 的平均值分别为 0.6 和 0.67 外，其他所有变量的 R 和 R ² 值均高于 0.89 和 0.80。所有变量的 RMSE 值也都在可接受范围内（表 3）。SDSM 在温度方面表现最好，而在 WS 方面表现相对较差。此外，还发现验证结果甚至比校准结果更好一些。

Table 3. Performance Evaluation of Statistical DownScaling Model During Calibration (1981–2005) and Validation (2006–2015) Periods in the Three-River Headwater Region
表 3.三河源头地区校准期（1981-2005 年）和验证期（2006-2015 年）统计降尺度模型的性能评估

	Calibration (1981–2005) 校准（1981-2005 年）			Validation (2006–2015) 验证（2006-2015 年）
	R	R2	RMSE	R	R2	RMSE
Precipitation 降水量
Maximum 最大	0.93	0.87	22.60	0.955	0.912	26.25
Minimum 最低	0.88	0.77	12.57	0.865	0.748	11.67
Mean 平均值	0.91	0.83	17.99	0.915	0.839	19.18
Maximum temperature 最高温度
Maximum 最大	0.99	0.99	1.91	0.995	0.990	1.10
Minimum 最低	0.97	0.94	0.68	0.988	0.977	0.62
Mean 平均值	0.98	0.96	1.28	0.993	0.986	0.78
Minimum temperature 最低温度
Maximum 最大	1.00	0.99	2.01	0.996	0.992	1.25
Minimum 最低	0.98	0.96	0.80	0.990	0.981	0.66
Mean 平均值	0.99	0.98	1.12	0.994	0.988	0.96
Relative humidity
Maximum 最大	0.95	0.90	6.66	0.962	0.925	5.29
Minimum 最低	0.78	0.60	3.16	0.881	0.776	2.93
Mean 平均值	0.89	0.80	5.01	0.924	0.854	4.58
Wind speed 风速
Maximum 最大	0.93	0.86	0.54	0.961	0.923	0.34
Minimum 最低	0.65	0.42	0.23	0.624	0.389	0.15
Mean 平均值	0.77	0.60	0.35	0.816	0.673	0.23

Additionally, to check the performance of GCMs, SDSM was forced with the GCM predictors to simulate climate variables for the historical period 1981–2015 and compared with the observations using the statistical indicators, as described in Table 4. R and R² for the three GCMs ranged between 0.64–0.99 and 0.41–0.98, while varied from 0.88 to 0.99 and 0.77–0.99 for the ensemble of GCMs (GCM-ENSM). RMSE values for PREC, Tmax/Tmin, RH, and WS were less than 25 mm/month, 2°C, 6%, and 0.33 m/s, respectively, for all GCMs and GCM-ENSM. Furthermore, an annual cycle of each variable was plotted for all GCMs and GCM-ENSM against the observations for a deep investigation of simulated results, as shown in Figure 7. In the case of PREC, NorESM well captured the peaks months, while other models followed well the low PREC months. Nonetheless, Tmax and Tmin were well simulated by all the GCMs. Table 4 and Figure 7 showed that all variables simulated by forcing GCM-ENSM data performed better than individual GCM. Therefore, GCM-ENSM was applied for further analysis in the TRHR, as an ensemble mean provides more dependable and robust estimations than an individual model (Tebaldi & Knutti, 2007).
此外，为了检验 GCM 的性能，使用 GCM 预测因子强制 SDSM 模拟 1981-2015 年历史时期的气候变量，并使用统计指标与观测结果进行比较，如表 4 所述。三个 GCM 的 R 和 R ² 在 0.64-0.99 和 0.41-0.98 之间，而 GCM 集合（GCM-ENSM）的 R 和 R ² 在 0.88-0.99 和 0.77-0.99 之间。所有 GCM 和 GCM-ENSM 的 PREC、Tmax/Tmin、RH 和 WS 的 RMSE 值分别小于 25 毫米/月、2°C、6% 和 0.33 米/秒。此外，还将所有 GCM 和 GCM-ENSM 的各变量年周期与观测结果进行了对比，以深入研究模拟结果，如图 7 所示。在 PREC 方面，NorESM 很好地捕捉到了峰值月份，而其他模式则很好地捕捉到了 PREC 低值月份。不过，所有 GCM 都很好地模拟了 Tmax 和 Tmin。表 4 和图 7 显示，通过 GCM-ENSM 数据强迫模拟的所有变量均优于单个 GCM。因此，GCM-ENSM 被用于 TRHR 的进一步分析，因为集合平均值比单个模型提供了更可靠、更稳健的估计（Tebaldi 和 Knutti，2007 年）。

Table 4. Evaluation of Global Climate Models (GCMs) and Ensemble for the Historical Period 1981–2015 in the Three-River Headwater Region
表 4.三江源地区 1981-2015 年历史时期全球气候模式（GCM）和集合的评估

Variable 可变	GCMs 全球监测模块	R	R2	RMSE
Precipitation 降水量	CanESM	0.83	0.69	23.8
	NorESM	0.91	0.83	17.6
	MPI-ESM	0.91	0.84	17.3
	Ensemble mean 集合平均数	0.93	0.86	16.35
Maximum temperature 最高温度	CanESM	0.96	0.92	1.82
	NorESM	0.97	0.94	1.75
	MPI-ESM	0.96	0.91	1.97
	Ensemble 合奏	0.98	0.97	1.18
Minimum temperature 最低温度	CanESM	0.98	0.97	1.6
	NorESM	0.99	0.98	1.6
	MPI-ESM	0.99	0.98	1.4
	Ensemble mean 集合平均数	0.99	0.99	1.01
Relative humidity 相对湿度	CanESM	0.87	0.76	5.75
	NorESM	0.88	0.77	5.30
	MPI-ESM	0.87	0.75	5.71
	Ensemble 合奏	0.94	0.89	3.78
Wind speed 风速	CanESM	0.64	0.41	0.33
	NorESM	0.74	0.55	0.27
	MPI-ESM	0.70	0.48	0.29
	Ensemble mean 集合平均数	0.88	0.77	0.20

4.2 HEC-HMS's Performance Evaluation
4.2 HEC-HMS 的性能评估

Calibration and validation Table 5 shows the Nash-Sutcliffe efficiency (E), coefficient of determination (R²), NRMSE, and percent bias (PBIAS) calculated from monthly streamflow at five hydrometric stations. The E and R², which are used to quantify how well a model can predict the outcome variable, ranged from 0.70 to 0.93, and the PBIAS values were less than 10% at all gauges both for calibration and validation, except at Jimai having PBIAS of 12% and 21% during validation period-1 and -2, respectively. According to Moriasi et al. (2015), the performance of a hydrological model is considered satisfactory if E > 0.50 and PBIAS ≤±25% and very good if E > 0.75 and PBIAS ≤±10%. NRMSE varied from 0.21 to 0.37, which was a little higher but still within the acceptable range, as it should be less than half of the SD of measured data (J. Singh et al., 2004). Additionally, simulated streamflow at Xiangda, Zhimenda, and Tangnaihai was plotted against the observations to analyze how well the model's simulations capture the monthly variations during each year, which are shown in Figure 8. The graphs showed that the model well captured all the components of observed hydrograph (e.g., baseflow, falling limb, rising limb, and time to peaks) except peaks. For example, at Xiangda, the model overestimated in 2008, 2009, and 2012, while underestimated during 2010, 2011, and 2013. On the whole, the statistical indicators and hydrographs showed that the model was well established to simulate streamflow in the region.
校准和验证 表 5 显示了根据五个水文站的月度流量计算得出的纳什-苏特克利夫效率 (E)、判定系数 (R ² )、NRMSE 和偏差百分比 (PBIAS)。E 和 R ² 用于量化模型预测结果变量的程度，范围在 0.70 到 0.93 之间，除吉迈站在验证期-1 和-2 的 PBIAS 分别为 12% 和 21% 外，校准和验证期间所有测站的 PBIAS 值均小于 10%。根据 Moriasi 等人（2015 年）的研究，如果 E > 0.50 且 PBIAS ≤±25% 则认为水文模型的性能令人满意，如果 E > 0.75 且 PBIAS ≤±10%，则认为水文模型的性能非常好。NRMSE 从 0.21 到 0.37 不等，略高但仍在可接受范围内，因为它应小于测量数据 SD 的一半（J. Singh 等，2004 年）。此外，还将向大、直门达、唐乃亥的模拟流量与观测数据进行了对比，以分析模型模拟对每年月度变化的捕捉程度，如图 8 所示。从图中可以看出，除峰值外，模型很好地捕捉了观测水文图的所有组成部分（如基流、下降沿、上升沿和达到峰值的时间）。例如，在湘大，模型在 2008 年、2009 年和 2012 年高估了水量，而在 2010 年、2011 年和 2013 年低估了水量。总体而言，统计指标和水文图显示，该模型模拟该地区河水流量的能力较强。

Table 5. Calibration and Validation of HEC-HMS in the Three-River Headwater Region
表 5.三河源头地区 HEC-HMS 的校准和验证

	Jimai 吉迈	Maqu 马库斯	Tangnaihai 唐乃亥	Zhimenda 日门达	Xiangda 湘大
Calibration (2006–2010) 校准（2006-2010 年）
E	0.74	0.88	0.87	0.85	0.89
R2	0.79	0.91	0.92	0.92	0.89
PBIAS (%)	4.65	3.53	6.14	9.10	−5.68 -5.68
NRMSE	0.37	0.25	0.24	0.36	0.24
Validation-1 (2011–2015) 验证-1（2011-2015 年）
E	0.77	0.86	0.93	0.85	0.91
R2	0.85	0.87	0.93	0.89	0.91
PBIAS (%)	12.8	3.29	−1.32 -1.32	8.62	6.49
NRMSE	0.31	0.29	0.21	0.36	0.26
Validation-2 (1981–2005) 验证-2（1981-2005 年）
E	0.70	0.86	0.86	0.89	0.84
R2	0.78	0.87	0.89	0.90	0.85
PBIAS (%)	21.20	−4.30 -4.30	−0.79 -0.79	1.34	−1.13 -1.13
NRMSE	0.37	0.29	0.31	0.32	0.31

4.2.1 Evaluation of Hydrological Components
4.2.1 水文成分评估

For the evaluation of HEC-HMS simulating SMC, TWS, SWE, and AET, the data was obtained from ESA-CCI-SM, the GRACE solution of UT-CSR, Advanced Microwave Scanning Radiometer-Earth Observing System, and TERACLIMATE products, respectively, described in Table 1. However, baseflow was evaluated with that separated by the RDF. Since GRACE-TWS is available in an anomaly format relative to the mean of 2004–2009, we also calculated water storage anomalies (HEC-TWSA) relative to 2004–2009, as in Yuan et al. (2018). Three statistical indicators (i.e., correlation coefficient [R], PBIAS, and RMSE) were used for the evaluation of these components, which were calculated from monthly time series and are described in Table 6. The R values ranged from 0.66 to 0.96 in all three basins for all the components, except in the HYaR for SWE where the R value was 0.33. The highest correlations were observed in the case of baseflow followed by AET. PBIAS varied between −13% and 20% in all three basins except for SWE in the HLaR and HYeR. On the whole, the best results were observed for baseflow followed by AET, SMC, and TWS. The SWE comparison with remote sensing data was not as good as compared to other components, which might be due to uncertainties exhibited in remote sensing data. It is required to evaluate SWE with other remote sensing products as well as observations.
在对模拟 SMC、TWS、SWE 和 AET 的 HEC-HMS 进行评估时，数据分别来自 ESA-CCI-SM、UT-CSR 的 GRACE 解决方案、高级微波扫描辐射计-地球观测系统和 TERACLIMATE 产品，详见表 1。不过，基流是用 RDF 分离出来的基流进行评估的。由于 GRACE-TWS 有相对于 2004-2009 年平均值的异常格式，我们也计算了相对于 2004-2009 年的蓄水异常（HEC-TWSA），如 Yuan 等（2018 年）所述。表 6 介绍了根据月度时间序列计算的三个统计指标（即相关系数 [R]、PBIAS 和 RMSE），用于评估这些分量。在所有三个流域中，所有成分的 R 值都在 0.66 到 0.96 之间，只有降水量的 HYaR 的 R 值为 0.33。基流的相关性最高，其次是 AET。在所有三个流域中，PBIAS 的变化范围在 -13% 到 20% 之间，但在 HLaR 和 HYeR 中的 SWE 除外。总体而言，基流的结果最好，其次是 AET、SMC 和 TWS。与其他成分相比，SWE 与遥感数据的比较结果并不理想，这可能是由于遥感数据的不确定性造成的。需要用其他遥感产品和观测数据对 SWE 进行评估。

Table 6. Evaluation of Different Hydrological Components in the Three-River Headwater Region
表 6.三河源头地区不同水文要素的评估

Hydrological components → 水文成分 →	SMC	SWE	Baseflow 基流	TWSC	AET
RRSD →	ESA-CCI-SM	AMSR-E	RDF	UT-CSR	TERRACLIMATE
Analysis period → 分析期→	2001–2010 2001-2010	2001–2010 2001-2010	1981–2015 1981-2015	2002–2015 2002-2015	2000–2010 2000-2010
The headwater of the Lancang River 澜沧江源头
R	0.72	0.75	0.95	0.70	0.89
RMSE	0.03	12.50	29.00	22.10	15.30
PBIAS (%)	4.50	−51.70 -51.70	18.10	−5.30 -5.30	0.10
The headwater of the Yellow River 黄河源头
R	0.72	0.73	0.96	0.75	0.85
RMSE	0.03	6.30	100.80	20.40	21.90
PBIAS (%)	6.70	−59.80 -59.80	7.40	9.10	−13.50 -13.50
The headwater of the Yangtze River 长江源头
R	0.80	0.33	0.96	0.66	0.86
RMSE	0.03	8.20	113.20	19.20	17.30
PBIAS (%)	11.50	−11.30	20.10	5.10	−3.50

• Note. RRSD, reanalysis and remote sensing data; DRF, Recursive Digital Filter; SMC, Soil moisture content; SWE, Snow water equivalent; TWSC, Terrestrial water storage changes; AET, Actual evapotranspiration.
  注。RRSD，再分析和遥感数据；DRF，递归数字滤波；SMC，土壤水分含量；SWE，雪水当量；TWSC，陆地储水变化；AET，实际蒸散量。

4.3 Projected Hydroclimatic Changes
4.3 预计的水文气象变化

4.3.1 Mean Annual Changes
4.3.1 年均变化

Table 7 describes the projected changes in climatic (i.e., Tmax, Tmin, PREC, RH, and WS) and hydrological (i.e., AET, streamflow, baseflow, surface flow, snow-melt flow, TWS, and SMC) elements in the NFP (2021–2060) and FFP (2061–2100) with respect to the BLP (1981–2020), simulated under SSP2-4.5 in the TRHR. Tmax (Tmin) were projected to increase by 1.3–1.5 (0.9–1.0)°C and 2.3–2.8 (1.4–2.0)°C in the NFP and FFP, respectively, with relatively high values in the HYeR. Similarly, positive changes, ranging from 2.3% to 8.1%, were determined in RH, with a maximum in the HLaR. WS was predicted to decrease in the HLaR and HYeR but increase in the HYaR. PREC, one of the most important water cycle components, showed an increase of 9.0%–21.6% in the NFP and 16.7%–40.3% in the FFP, with a maximum increase in the HYeR. This projected increase in PREC caused positive changes in all the hydrological components. AET, streamflow, baseflow, surface flow, snow-melt flow, TWS, and SMC were predicted to increase by 6.3%–30.6%, 9.0%–95.0%, 7.9%–58%, 24.0%–426.0%, 0.1%–12.5%, and 0.5%–9.7%, respectively, under both scenarios.
表 7 描述了 TRHR 中 SSP2-4.5 模拟的 NFP（2021-2060 年）和 FFP（2061-2100 年）相对于 BLP（1981-2020 年）的气候（即 Tmax、Tmin、PREC、RH 和 WS）和水文（即 AET、溪流、基流、地表流、融雪流、TWS 和 SMC）要素的预计变化。预计在 NFP 和 FFP 中，Tmax（Tmin）将分别增加 1.3-1.5 (0.9-1.0)°C 和 2.3-2.8 (1.4-2.0)°C，在 HYeR 中的数值相对较高。同样，相对湿度也出现了正变化，变化范围从 2.3% 到 8.1%，在 HLaR 中变化最大。据预测，WS 在 HLaR 和 HYeR 将下降，但在 HYaR 将上升。PREC 是水循环最重要的组成部分之一，在 NFP 和 FFP 中的增幅分别为 9.0%-21.6% 和 16.7%-40.3%，在 HYeR 中增幅最大。预计 PREC 的增加会导致所有水文成分发生积极变化。在这两种情景下，预计 AET、溪流、基流、地表流、融雪流、TWS 和 SMC 分别增加 6.3%-30.6%、9.0%-95.0%、7.9%-58%、24.0%-426.0%、0.1%-12.5% 和 0.5%-9.7%。

Table 7. Projected Hydro-Meteorological Changes in 2021–2060 and 2061–2100 Relative to 1981–2020, Under SSP2-4.5 in the Three-River Headwater Region
表 7.在 SSP2-4.5 条件下，三江源头地区 2021-2060 年和 2061-2100 年相对于 1981-2020 年的水文气象变化预测值

	Headwater of Lancang River 澜沧江源头			Headwater of Yangtze 长江源头			Headwater of Yellow 黄河源头
	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100
	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)
Tmax 最大值	10.2	1.3	2.3	7.0	1.5	2.6	8.3	1.5	2.8
Tmin	−3.4 -3.4	0.9	1.5	−7.2 -7.2	0.7	1.4	−7.2 -7.2	1.0	2.0
RH	54.5	4.5	8.1	54.7	3.6	5.8	68.9	2.3	4.1
WS	1.6	−3.7 -3.7	−6.7 -6.7	2.4	6.2	4.2	2.6	−0.9 -0.9	−1.2 -1.2
PREC	596	9.0	16.7	449	17.2	30.5	615	21.6	40.3
AET	314	6.3	11.2	383	10.9	19.9	521	17.4	30.6
QSIM	290	9.0	19.3	71	45.6	80.6	94	46.4	95.0
QB	270	7.9	14.8	64	38.0	58.0	84	34.6	56.4
QS	19	24.6	82.4	7	118.2	299.1	10	147.5	425.9
QSNOW	61	5.2	11.6	3	66.0	31.1	3	66.2	111.5
TWS	135	0.1	2.1	110	3.8	6.2	149	8.6	12.5
SM	103	0.5	1.7	96	2.1	3.0	121	5.7	9.7

• Note. Tmax, maximum temperature; Tmin, minimum temperature; RH, Relative humidity; WS, wind speed; PREC, precipitation; AET, actual evapotranspiration; Q_SIM, simulated streamflow; Q_B, baseflow; Q_S, surface flow (direct flow); Q_SNOW, snowmelt flow; TWS, terrestrial water storage; SM, soil moisture.
  注。Tmax，最高气温；Tmin，最低气温；RH，相对湿度；WS，风速；PREC，降水量；AET，实际蒸散量；Q _SIM ，模拟溪流；Q _B ，基流；Q _S ，地表流（直接流）；Q _SNOW ，融雪流；TWS，陆地蓄水量；SM，土壤水分。

Under SSP5-5.8, the potential changes for hydro-climatic elements in the NFP and FFP relative to the BLP are described in Table 8. Almost all the changes (positive or negative) were quite similar to that under SSP2-4.5 but the magnitudes were much higher than that under SSP2-4.5. For example, Tmax was predicted to rise by 1.7–2.0°C and 4.0–4.6°C in the NFP and FFP, respectively, which was almost 1.3–2.0 times higher than that under SSP2-4.5. Similarly, PREC was increased by 15.8%–26.8% and 43.0%–86.8% in the NFP and FFP, respectively, which was almost, 1.4–2.2 times higher than that under SSP2-4.5.
表 8 介绍了 SSP5-5.8 条件下，相对于最惠国待遇方案，国家联络点和联邦联络点水文气象要素的潜在变化。几乎所有的变化（正或负）都与 SSP2-4.5 条件下的变化非常相似，但变化幅度要比 SSP2-4.5 条件下的变化大得多。例如，预测 NFP 和 FFP 的 Tmax 分别升高 1.7-2.0°C 和 4.0-4.6°C，几乎是 SSP2-4.5 的 1.3-2.0 倍。同样，在 NFP 和 FFP 中，PREC 分别增加了 15.8%-26.8%和 43.0%-86.8%，几乎是 SSP2-4.5 的 1.4-2.2 倍。

Table 8. Projected Hydro-Meteorological Changes in 2021–2060 and 2061–2100 Relative to 1981–2020 Under SSP5-5.8 in the Three-River Headwater Region
表 8.在 SSP5-5.8 条件下，三江源头地区 2021-2060 年和 2061-2100 年相对于 1981-2020 年的水文气象变化预测值

	Headwater of Lancang River 澜沧江源头			Headwater of Yangtze 长江源头			Headwater of Yellow 黄河源头
	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100	1981–2020 1981-2020	2021–2060 2021-2060	2061–2100 2061-2100
	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)	Abs (°C/mm)	Δ (°C/%)	Δ (°C/%)
Tmax 最大值	10.2	1.7	4.0	7.0	1.8	4.5	8.3	1.9	4.6
Tmin	−3.4 -3.4	1.1	2.7	−7.2 -7.2	1.0	2.8	−7.2 -7.2	1.3	3.6
RH	54.5	6.0	16.4	54.7	4.4	12.1	68.9	2.8	8.1
WS	1.6	−4.7 -4.7	−11.9 -11.9	2.4	4.9	−0.2 -0.2	2.6	−1.3 -1.3	−2.7 -2.7
PREC	596	15.8	43.1	449	24.8	65.9	612	26.8	86.8
AET	314	9.3	24.2	389.4	14.4	37.5	514	28.8	73.2
QSIM	282	22.6	59.6	70.3	64.6	198.5	98	16.2	157.3
QB	264	18.0	29.6	63.9	48.3	102.8	79	18.7	68.1
QS	19	87.7	480.7	6.4	228.1	1,053.7	18	5.7	540.7
QSNOW	60	14.8	34.3	3.5	76.5	17.0	3	78.0	128.9
TWS	134	3.1	7.7	116	5.4	13.5	147	10.0	21.7
SM	102	2.4	5.7	96	3.0	7.3	121	6.4	15.7

• Note. Tmax, maximum temperature; Tmin, minimum temperature; RH, Relative humidity; WS, wind speed; PREC, precipitation; AET, actual evapotranspiration; Q_SIM, simulated streamflow; Q_B, baseflow; Q_S, surface flow (direct flow); Q_SNOW, snowmelt flow; TWS, terrestrial water storage; SM, soil moisture.
  注。Tmax，最高气温；Tmin，最低气温；RH，相对湿度；WS，风速；PREC，降水量；AET，实际蒸散量；Q _SIM ，模拟溪流；Q _B ，基流；Q _S ，地表流（直接流）；Q _SNOW ，融雪流；TWS，陆地蓄水量；SM，土壤水分。

Although the increased temperature accelerated AET, the increased PREC dominated the region, causing an increase in all the hydrological components. It was observed that the hydro-climatic changes were mostly maximum in the HYeR followed by HYaR, which was mainly due to higher changes in temperature and PREC in the HYeR. The highest rise was observed in surface flow (direct flow) under both scenarios, indicating more floods in the future. All hydrological components showed higher changes in the FFP as compared to the NFP, except snow-melt flow in the HYaR.
虽然气温升高加速了 AET 的增加，但 PREC 的增加在该区域占主导地位，导致所有水文要素的增加。据观察，水文气候的变化主要在 HYeR 中最大，其次是 HYaR，这主要是由于 HYeR 中温度和 PREC 变化较大。在这两种情景下，地表流量（直接流）的上升幅度最大，这表明未来会有更多的洪水。与 NFP 相比，除 HYaR 中的融雪流量外，FFP 中所有水文成分的变化都更大。

4.3.2 Annual Cycle Changes
4.3.2 年周期变化

Figures 9 and 10 show the annual cycles of hydro-climatic components for the BLP, NFP, and FFP under SSP2-4.5 and SSP5-8.5, respectively. Instead of individual headwaters, these cycles were plotted for the whole TRHR, by taking the weighted mean. On the whole, almost all variables showed profound intensification in the NFP and FFP, under both scenarios. The changes were even much higher in the FFP and under SSP5-8.5. Climate variables, that is, Tmax, Tmin, and RH displayed an increase in all months, while WS showed a decrease, especially under SSP5-8.5. On the other hand, the hydrological components such as PREC, streamflow (runoff), and AET were projected to increase mainly during the melting and rainy season (April–October) relative to the BLP. For example, PREC showed a significant increase from May to October, with peaks shifting toward August and September (Figures 9e and 10e). Consequently, the PREC-dependent hydrological components such as streamflow, baseflow, SMC, and TWS exhibited similar patterns to PREC. Streamflow and baseflow also displayed higher values in the winter and spring seasons, especially under SSP5-8.5 and in the FFP, though PREC did not. This can be mainly due to increased temperature in the FFP, causing early snow/ice melting. This melting effect can also be seen in Figures 9j and 10j, where melting is starting in February instead of April. A profoundly high annual cycle in direct flow was projected under both scenarios, especially in the FFP, an indication of high floods in the future.
图 9 和图 10 分别显示了 SSP2-4.5 和 SSP5-8.5 条件下 BLP、NFP 和 FFP 的水文气候成分年周期。这些周期是通过加权平均值绘制的整个 TRHR 的周期，而不是单个水源地的周期。总体而言，在这两种情况下，几乎所有变量都显示出 NFP 和 FFP 的严重加剧。在全境平均温度和 SSP5-8.5 条件下，变化甚至更大。气候变量，即最高气温、最低气温和相对湿度在所有月份都有所上升，而湿度则有所下降，特别是在 SSP5-8.5 条件下。另一方面，与 BLP 相比，PREC、溪流（径流）和 AET 等水文要素预计主要在融化和雨季（4 月至 10 月）期间增加。例如，PREC 在 5 月至 10 月期间显著增加，峰值转向 8 月和 9 月（图 9e 和 10e）。因此，与 PREC 相关的水文成分，如径流量、基流、SMC 和 TWS 也表现出与 PREC 相似的模式。在冬季和春季，尤其是在 SSP5-8.5 条件下和在 "粮食丰收季节 "中，河水流量和基流也显示出较高的值，而 PREC 则没有。这可能主要是由于全境平均温度升高，导致冰雪提前融化。从图 9j 和图 10j 中也可以看到这种融化效应，融化从二月而不是四月开始。在这两种情景下，预计直接流量的年周期都很高，特别是在森林覆盖区，这表明未来洪水会很高。

4.3.3 Evolution of Hydro-Climatic Elements
4.3.3 水文气候要素的演变

To analyze the temporal changes in hydro-climatic elements over the whole study period (1981–2100), annual time series plots are shown in Figure 11 along with the rate of change for the whole period, which were estimated by linear regression. On the whole, all elements showed continuous increasing trends under both scenarios, except WS. However, ROCs were higher under SSP5-8.5 than that under SSP2-4.5, even extremely high after 2050. It was observed that the elements started stabilizing after 2050 under SSP2-4.5, while exaggerated under SSP5-8.5. For example, Tmax (Tmin) was estimated to increase at the rate of 0.29–0.35 (0.19–0.24)°C/10 year and 0.5–0.57 (0.34–0.46)°C/10 year under SSP2-4.5 and SSP5-8.5, respectively, showing almost double figures under SSP5-8.5 (Figures 11a and 11b). Similarly, the ROCs of PREC under SSP5-8.5 were almost two times higher than that under SSP2-4.5, ranging from 11 to 31 mm/10 year and 33–67 mm/10 year, respectively (Figure 11e). Nonetheless, the ROCs of AET were smaller than PREC, ranging from 4 to 20 mm/10 year under SSP2-4.5 and 10–47 mm/10 year under SSP5-8.5 (Figure 11f). Since the ROCs of PREC dominated the AET, this can be the main reason for an increase in all other hydrological components such as streamflow, surface flow, and SMC. For example, streamflow will increase at the rate of 7–11 mm/10 year and 18–24 mm/10 year under SSP2-4.5 and SSP5-8.5, respectively (Figure 11g). Although all hydrological components showed high ROCs in the last decade (2091–2100) relative to the rest period (1981–2090) under SSP5-8.5, surface flow displayed an unprecedented increase during this period. So, the last decade will face a lot of flooding in the region. It was also noticed that ROCs were higher in the HYeR than the other two basins under both scenarios in the case of almost all hydroclimatic elements except melt-flow.
为了分析整个研究期间（1981-2100 年）水文气候要素的时间变化，图 11 显示了年度时间序列图和整个研究期间的变化率，这些变化率是通过线性回归估算得出的。总体而言，除 WS 外，所有要素在两种情景下均呈持续增长趋势。但是，SSP5-8.5 条件下的 ROCs 要高于 SSP2-4.5 条件下的 ROCs，甚至在 2050 年后达到极高值。据观察，在 SSP2-4.5 条件下，各要素在 2050 年后开始趋于稳定，而在 SSP5-8.5 条件下则有所夸大。例如，在 SSP2-4.5 和 SSP5-8.5 条件下，Tmax（Tmin）估计分别以 0.29-0.35 (0.19-0.24)℃/10 年和 0.5-0.57 (0.34-0.46)℃/10 年的速度增加，而在 SSP5-8.5 条件下几乎增加了一倍（图 11a 和 11b）。同样，SSP5-8.5 条件下的 PREC ROCs 也比 SSP2-4.5 条件下的 ROCs 高出近两倍，分别为 11 至 31 毫米/10 年和 33 至 67 毫米/10 年（图 11e）。然而，AET 的 ROC 小于 PREC，在 SSP2-4.5 条件下为 4 至 20 毫米/10 年，在 SSP5-8.5 条件下为 10 至 47 毫米/10 年（图 11f）。由于 PREC 的 ROCs 在 AET 中占主导地位，这可能是所有其他水文成分（如溪流、地表流量和 SMC）增加的主要原因。例如，在 SSP2-4.5 和 SSP5-8.5 条件下，溪流将分别以 7-11 毫米/10 年和 18-24 毫米/10 年的速度增加（图 11g）。虽然在 SSP5-8.5 条件下，最后十年（2091-2100 年）所有水文要素的 ROC 值都比其余时期（1981-2090 年）高，但地表流量在此期间出现了前所未有的增长。因此，最后十年该地区将面临大量洪水。 同时还注意到，在两种情景下，几乎所有水文气候要素（融流除外）的 ROCs 在 HYeR 都高于其他两个盆地。

5 Discussion 5 讨论

5.1 Hotter and Wetter Future
5.1 未来更热、更潮湿

Hydroclimatic projections of a region are critical to cope with climate change and long-term water resources planning and management in a better way. These kinds of scientific-based predictions offer a great source of information to help policy-makers, regional communities as well as governments in planning and adapting the projected hydrological changes such as changes in water availability. In this study, a statistical downscaling method along with a hydrological model was applied to assess the hydrological responses to changing climate under two scenarios for the future relative to the BLP. The results showed that all hydrological components such as streamflow, baseflow, surface flow, SMC, and AET responded positively (increase) to the projected climate such as increased temperature, PREC, and humidity in the TRHR. Most previous studies exhibited similar kind of results in and around the study area, increased streamflow (runoff) in the TRHR such as by F. Su et al. (2016) over the TRHR (5%–20%, 2011–2070, 3-RCPs); T. Wang et al. (2022) over the HYaR (1–7 mm, 2021–2070, 3-SSPs [1,2,5]); Q. Zhao et al. (2019) over the TRHR (6%–21%, 2050s, RCP2.6 and 4.5), Y. Zhao et al. (2022) over the Yalong River basin (12–25 mm/10 year, SSP2 and 5); Lu et al. (2018) over the HYeR (3%) and HYaR (5%). However, some studies showed contradicting results. For example, Immerzeel et al. (2010) explored a 5.2% decrease in the HYaR in 2046–2065 under A1B and a 16%–20% decreased streamflow by Hu et al. (2022) over the HYeR. The main reason for contradicting results might be due to using different downscaling methods, hydrological models, input data sets, scenarios, and analysis periods for both baseline and future.
一个地区的水文气候预测对于更好地应对气候变化和长期水资源规划与管理至关重要。这些以科学为基础的预测提供了大量信息，有助于政策制定者、地区社区和政府规划和适应预计的水文变化，如供水量的变化。在这项研究中，采用了一种统计降尺度方法和一个水文模型，以评估未来相对于 BLP 的两种情景下水文对气候变化的反应。结果显示，TRHR 的所有水文成分，如溪流、基流、地表水流、SMC 和 AET 都对预测的气候，如温度、PREC 和湿度的增加做出了积极的响应（增加）。大多数先前的研究都在研究区域及其周边地区显示出类似的结果，即 TRHR 地区的河川流量（径流）增加，如 F. Su 等人（2016 年）对 TRHR 地区的研究（5%-20%，2011-2070 年，3-RCPs）；T. Wang 等人（2022 年）对 HYRAM 地区的研究（5%-20%，2011-2070 年，3-RCPs）。(2022 年）；Q. Zhao 等（2019 年），TRHR（6%-21%，2050 年代，RCP2.6 和 4.5）；Y. Zhao 等（2022 年），雅砻江流域（12-25 毫米/10 年，SSP2 和 5）；Lu 等（2018 年），HYeR（3%）和 HYaR（5%）。然而，一些研究结果却相互矛盾。例如，Immerzeel 等人（2010 年）探讨了 A1B 条件下 2046-2065 年 HYaR 减少 5.2%，而 Hu 等人（2022 年）探讨了 HYeR 减少 16%-20%。结果相互矛盾的主要原因可能是对基线和未来使用了不同的降尺度方法、水文模型、输入数据集、情景和分析期。

It is quite obvious that increased flows (e.g., streamflow, surface flow, and melt-flow) in the region are mainly due to projected increased PREC because it is the primary factor affecting the flow regimes (T. Zhang et al., 2018). IPCC (2021) showed positive changes in PREC not only on the global level but also over China and the TRHR. Over China, they showed a 10%–40% increase under 1.5–4.0°C of global warming scenarios. The PREC projections in the region are also in-line with the previous studies such as F. Su et al. (2016), T. Wang et al. (2022), and Q. Zhao et al. (2019), however, with different increasing magnitudes in the TRHR (ranging from 10% to 40% increase at the end of twenty-first century under different RCPs). In addition, Lu et al. (2018) showed an increase in PREC up to 70% in some parts of the TP in 2041–2060. However, the projected increase in PREC in the TRHR is much higher than the increased global land PREC, which is about 1.5%–8% and 1%–13% under SSP2-4.5 and SSP5-8.5 in 2081–2100. It is also reported that every 1°C increase in global warming will increase PREC by about 7% (IPCC, 2021). This means the TRHR will be wetter and wetter in the twenty-first century because temperature in the region will increase by 1.0–1.5 (1–1.9)°C and 2.0–2.8 (2.7–4.6)°C in 2021–2060 and 2061–2100 relative to 1981–2020 under SSP2-4.5 (SSP5-8.5), respectively, which was highly expected in the region because it is well established that temperature will increase in the future on global as well as region scales. However, the increasing magnitude could vary from region to region. For example, similar increasing values have been explored in the previous studies conducted in and/or around the TRSR such as by Gu et al. (2018) over the Yangtze River basin above Yichang (1.81–2.26°C, 2021–2050, RCP4.5–RCP8.5); Hu et al. (2022) over HYeR above Tangnaihai (1.4–4.5°C, 2061–2100, RCP4.5 and RCP8.5), which are little higher than this study; Lu et al. (2018) over TP (1–4°C, 2041–2060, RCP2.6–8.5); Q. Zhao et al. (2019) over TP (1.9–3.5°C, 2090–2100, RCP2.6–4.5). Similarly, an increase in global mean temperature of 1–5.7°C has been reported by IPCC (2021) under SSP1-1.9–SSP5-8.5 in 2081–2100 relative to the pre-industrial era. This is mainly due to the present and projected increasing concentration of greenhouse gases in the atmosphere due to burning fossil fuel, deforestation, agricultural activities, etc. (J. Chu et al., 2010). It was also noticed that the studies that estimated a higher increasing percentage of ET than PREC exhibited decreasing streamflow such as Hu et al. (2022). This means the TRHR will be wetter and hotter in the future, with much more rain and snow, and a higher risk of flooding because surface flow was projected to be much higher, even up to a 500%–1,000% increase relative to the BLP. To explore the reason for this unprecedented increase in surface flow, some extreme precipitation indices, that is, P-1 day (maximum annual daily precipitation), P-5 day (maximum annual 5-daily precipitation), P20 (annual count of days when precipitation was ≥20 mm), P90P (90th percentile) and SD, were calculated for the BLP, NFP, and FFP, as in Vu et al. (2019). It was found that all extreme indices showed a profound increase in the FFP relative to the BLP. For example, P-1 day, P-5 day, P20, P90P, and SD increased by 40%, 52%, 1,200%, 38%, and 35%, respectively, in the FFP. Figure 12 shows that the extreme events will increase at a faster rate after 2070 in the FFP. Nonetheless, in the NFP, the precipitation extreme events were even less than that in the BLP. Increases in other hydrological components such as increased baseflow, surface flow, snowmelt flow, SMC, and TWS are mainly due to increased annual PREC and increased extreme events because it is the primary source of input for all these components. Despite some uncertainties, the results of the present study can be more reliable than the previous study because we calibrated and validated the SDSM, with high satisfaction. However, previous studies just corrected the GCM's outputs using bias correction methods.
很明显，该地区流量（如溪流、地表流和融流）的增加主要是由于预测的 PREC 增加，因为它是影响流量机制的主要因素（T. Zhang 等人，2018 年）。IPCC （2021 年）显示，PREC 不仅在全球范围内发生了积极变化，而且在中国和 TRHR 上也发生了积极变化。在全球变暖 1.5-4.0°C 的情景下，中国的 PREC 增长了 10%-40%。该区域的 PREC 预测也与之前的研究一致，如 F. Su 等人（2016 年）、T. Wang 等人（2022 年）和 Q. Zhao 等人（2019 年），但 TRHR 的增加幅度不同（在不同 RCPs 下，21 世纪末的增加幅度从 10%到 40% 不等）。此外，Lu 等人（2018 年）的研究表明，2041-2060 年，在热带降雨带的某些地区，PREC 的增幅高达 70%。然而，TRHR 的 PREC 预计增幅远高于全球陆地 PREC 的增幅，在 2081-2100 年 SSP2-4.5 和 SSP5-8.5 条件下，全球陆地 PREC 的增幅分别约为 1.5%-8% 和 1%-13%。另据报道，全球变暖每升高 1℃，PREC 将增加约 7%（IPCC，2021 年）。这意味着 TRHR 在二十一世纪将越来越湿润，因为在 SSP2-4.5（SSP5-8.5）条件下，2021-2060 年和 2061-2100 年该地区的气温将分别比 1981-2020 年上升 1.0-1.5 (1-1.9)°C 和 2.0-2.8 (2.7-4.6)° C。不过，不同地区的上升幅度可能会有所不同。例如，Gu 等人（2018 年）在宜昌以上长江流域（1.81-2.26°C，2021-2050 年，RCP4.5-RCP8.5)；Hu 等( 2022 年) 对唐乃亥以上 HYeR（1.4-4.5°C，2061-2100 年，RCP4.5 和 RCP8.5），略高于本研究；Lu 等（2018 年）在热带潮湿地区（1-4℃，2041-2060 年，RCP2.6-8.5）；Q. Zhao 等（2019 年）在热带潮湿地区（1.9-3.5℃，2090-2100 年，RCP2.6-4.5）。同样，IPCC（2021 年）报告称，在 SSP1-1.9-SSP5-8.5 条件下，2081-2100 年全球平均气温与工业化前相比将上升 1-5.7°C。这主要是由于燃烧化石燃料、砍伐森林、农业活动等导致大气中温室气体浓度不断增加（J. Chu 等人，2010 年）。研究还注意到，与 PREC 相比，那些估计蒸散发增加百分比较高的研究显示出河水流量的减少，如 Hu 等人（2022 年）。这意味着 TRHR 未来将更潮湿、更炎热，雨雪量更大，洪水风险更高，因为地表流量预计将大大增加，甚至比 BLP 增加 500%-1,000% 。为了探究地表流量空前增加的原因，与 Vu 等人（2019 年）一样，计算了 BLP、NFP 和 FFP 的一些极端降水指数，即 P-1（年最大日降水量）、P-5（年最大 5 日降水量）、P20（年降水量≥20 毫米的天数）、P90P（第 90 百分位数）和 SD。结果发现，相对于 BLP，FFP 的所有极端指数都有显著增加。例如，在 FFP 中，P-1 天、P-5 天、P20、P90P 和 SD 分别增加了 40%、52%、1,200%、38% 和 35%。图 12 显示，2070 年后，在粮食计划署，极端事件将以更快的速度增加。 然而，在国家边界点，极端降水事件甚至少于 BLP。其他水文成分的增加，如基流、地表流、融雪流、SMC 和 TWS 的增加，主要是由于年降水量增加和极端事件增加，因为它是所有这些成分的主要输入源。尽管存在一些不确定性，但由于我们对 SDSM 进行了校核和验证，因此本研究的结果比以前的研究更可靠，满意度很高。然而，以前的研究只是使用偏差校正方法校正了 GCM 的输出。

6 Conclusions 6 结论

Overall, the research provides valuable insights into the future hydrological dynamics of the TRHR, with implications for environmental sustainability, socio-economic development, and climate change adaptation in the region. The projected future changes in hydrological processes can be applied by policymakers, planners, and stakeholders for proactive adaptation strategies. This could include investments in water infrastructure, land use planning, ecosystem restoration, and community resilience-building initiatives to mitigate potential risks.
总之，这项研究为了解 TRHR 未来的水文动态提供了宝贵的见解，对该地区的环境可持续性、社会经济发展和气候变化适应具有重要意义。政策制定者、规划者和利益相关者可以利用对未来水文过程变化的预测，制定积极的适应战略。这可能包括对水利基础设施、土地利用规划、生态系统恢复和社区复原力建设举措进行投资，以减轻潜在风险。