Multiscale variability of streamflow in the Three Rivers Headwater Region, China, and links to large-scale atmospheric circulation indices

Streamflow is a component of fresh water resources that is vital to both human societies and natural ecosystems. Streamflows of major rivers are the product of integrated processes of changes in the regional atmosphere, hydrosphere, pedosphere, and cryosphere (DeBeer et al. 2016), and are greatly influenced by large-scale atmospheric circulation that varies on multiple time scales (Kingston et al. 2013; Li et al. 2019). In recent decades, researchers have paid closer attention to links between large-scale atmospheric circulation and streamflow in river basins. Such research could help provide explanations for regional streamflow change patterns and strengthen the understanding of the water cycle and hydrological processes (Cuo et al. 2014; Scott et al. 2019). These research topics are therefore essential for regional water resource management and protection (Ionita 2015; Tamaddun et al. 2016).

The Three Rivers Headwater Region (TRHR) is the source region of the Yellow, Yangtze, and Lancang-Mekong rivers; thus, it plays a key role in regulating the regional climate and water balance of East Asia (Yang et al. 2015). Over the last few decades, the TRHR has experienced evident climate changes (Liang et al. 2013; Tong et al. 2014), which have altered regional atmospheric and hydrological cycles (Immerzeel et al. 2010). Climate change and the streamflow changes accordingly induced in the TRHR directly affect the lives of people and animals that depend on the rivers originating from the TRHR. Therefore, understanding a streamflow change in the context of change in large-scale atmospheric circulation over the TRHR is essential for policymakers and water managers developing sustainable water resource strategies.

Numerous studies have been conducted on the changes in and correlations between hydrometeorological factors in the TRHR. Various statistical methods have been used to analyze the trend and periodic variations of TRHR hydrometeorological factors (Qian et al. 2014; Mao et al. 2016). Streamflow changes in the TRHR exhibit patterns similar to those of precipitation and temperature. The correlation coefficients between streamflow and precipitation are positive and larger than those between streamflow and temperature in the TRHR (Cuo et al. 2014). Correlation analysis and water balance model results were also used to identify the key factor influencing streamflow in the TRHR (Zhang et al. 2013). To improve understanding of the effects of climate change on streamflow, many efforts have been devoted to analyzing the relationships between large-scale atmospheric circulation systems and meteorological factors (Liu et al. 2012; Li et al. 2013). In winter, changes in precipitation in the TRHR can be attributed to changes in the North Atlantic Oscillation (NAO), Arctic Oscillation (AO), and South Asia monsoon (SAM). In summer, changes in precipitation are weakly related to these indices (Cuo et al. 2013). These studies cannot clearly represent the effects of large-scale atmospheric circulation systems on streamflow in this region (Liang et al. 2013). In addition, the effects of large-scale atmospheric circulation on streamflow are inconsistent among regions. Details of the changes from a holistic perspective are lacking for the three subregions of the TRHR (Mao et al. 2016). Therefore, the periodic behaviors of streamflow in the TRHR and their links to large-scale atmospheric circulation are still poorly understood. Research topics that clarify the links between large-scale atmospheric circulation and streamflow have yet to be addressed (Cuo et al. 2014; Yang et al. 2014).

In view of the mentioned shortcomings in the relevant literature, an ensemble empirical mode decomposition method is employed to analyze the periodical characteristics of TRHR streamflow. The relationships between streamflow and large-scale atmospheric circulation systems are investigated using cross-wavelet and wavelet coherence analysis methods on the basis of large-scale atmospheric circulation indices.

The TRHR is the source region of the Yangtze River, the Yellow River, and the Lancang-Mekong River. It is therefore known as the ‘water tower’ of China and Asia (Fang 2013). The TRHR is located in the core of the Tibetan Plateau and covers an area of 3.57 × 10⁵km², with an average altitude of 4,500 m above sea level. It provides 20% of the water volume of the Yangtze River, 49% of that of the Yellow River, and 15% of that of the Lancang River (Mao et al. 2016). It is considered to have a cold-dry climate, and its mean annual temperature ranges between −5.4 and 4.2 °C. Mean annual evaporation ranges between 730 and 1,700 mm (Yi et al. 2012). Precipitation in the TRHR is not equally distributed throughout the year, and roughly 86% of the total precipitation occurs during the rainy season (May–September).

The Zhimenda, Tangnaihai, and Changdu hydrological stations are national-level stations located at the outlets of the source regions of the Yangtze, Yellow, and Lancang rivers (Table 1; Zhang et al. 2013). In this research, streamflow data of the three stations were used to characterize streamflow variability in the three subregions of the TRHR (Figure 1).

Monthly streamflow data were provided by the Hydrology Bureau of Qinghai Province for the Zhimenda and Changdu hydrological stations and by the hydrological yearbook of China for the Tangnaihai hydrological station. Streamflow data from the three stations were quality-controlled and continuous from 1960 to 2012.

To clarify the relationships between large-scale atmospheric circulation systems and streamflow in the TRHR, six crucial and relatively independent large-scale atmospheric circulation indices, namely the AO index, the NAO index, the Southern Oscillation Index (SOI), the Tibetan Plateau Monsoon (TPM) index, SAM index, and the East Asian Monsoon (EAM) index, were considered in this study. The AO is a large-scale mode of climate variability and is also referred to as the northern hemisphere annular mode. The NAO is the dominant mode of winter climate variability in the North Atlantic region ranging from central North America to Europe and into much of Northern Asia. The AO and NAO indices were collected from https://www.cpc.ncep.noaa.gov/products/precip/CWlink/daily_ao_index/ao.shtml and https://www.cpc.ncep.noaa.gov/products/precip/CWlink/pna/nao.shtml, respectively. El Niño–Southern Oscillation (ENSO) is a naturally occurring phenomenon that involves fluctuating ocean temperatures in the equatorial Pacific, and the SOI is usually used to characterize the intensity of an ENSO event (Lal et al. 2013; McAfee & Wise 2016). SOI is a standardized index measure of large-scale fluctuations in air pressure occurring between the western and eastern tropical Pacific during El Niño and La Niña episodes. Therefore, the SOI was employed to characterize the intensity of an ENSO event in this research. The SOI indices were collected from http://www.cpc.ncep.noaa.gov/data/indices/soi. The TPM indices were produced from National Centers for Environmental Prediction (NCEP) reanalysis data (600 hPa geopotential height) by adopting the plateau monsoon index formula defined by Tang et al. (1984). The EAM indices were calculated from NCEP reanalysis data (850 hPa meridional wind averaged over 20–30°N and 110–130°E) by applying the formula defined by Li & Zeng (2005). The SAM indices were calculated based on the formula defined by Goswami et al. (1999); the differences in values between 850 and 200 hPa zonal wind were averaged over 10–30°N and 70–110°E. The NCEP reanalysis data were collected from http://www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis.pressure.html.

Ensemble empirical mode decomposition (EEMD) is a noise-assisted data analysis method (Wu & Huang 2009). EEMD has proven to be versatile in analyzing periodicity and trends in hydrometeorology time series (Sang et al. 2012; Rabbani et al. 2017). In the present study, EEMD is used to investigate the frequencies and trends of streamflow in three subregions of the TRHR.

In the EEMD method, the first step is adding a white noise series to the investigated signal x(t); second, the noise-added signal xm(t) is decomposed into a sum of intrinsic oscillatory components, each component an intrinsic mode function (IMF); third, these steps are repeated with new randomly generated noise series; finally, the means of corresponding IMFs of the decompositions are obtained as final IMFs. IMFs have different frequency components, based on which the trend, periodic components, and noise in the investigated signal can be identified and separated. The procedure is depicted in Figure 2.

Wavelet analysis is a powerful method for analyzing variability modes within a time series in finite domains (Torrence & Compo 1998). It has been widely applied to analyze variability in hydroclimatic processes at multitime scales (Sang 2013). The wavelet analysis method (cross-wavelet and wavelet coherence analyses) was therefore performed on the hydroclimatic time series of the TRHR.

Edge effects of the cross-wavelet and wavelet coherence spectra are excluded by using the method proposed by Grinsted et al. (2004). Statistically significant levels of the cross-wavelet and wavelet coherence spectra were implemented using Monte Carlo methods with a red noise (Torrence & Compo 1998).

Figure 3(a) presents EEMD analysis results for the annual streamflow time series at the Zhimenda hydrological station. Four IMFs with different periods were identified from the Zhimenda streamflow series during 1960–2012. IMF1 identified oscillations with an average period of approximately 4 years. IMF2 had an average period of approximately 7 years. IMF3 was an approximately 14-year cycle. IMF4 identified oscillations with an average period of approximately 36 years. The final residual indicated a clear increasing trend in the Zhimenda streamflow series.

The results of the significance tests for the IMFs, reflecting the importance of each component, are presented in Figure 3(b). The square points indicate the energy magnitude with a certain mean period for each IMF. If the square point is located above a line indicating a certain significance level, then the hypothesis that the IMF is induced by white noise is rejected at the corresponding significance level. Figure 2(b) shows that IMF3 in the Zhimenda streamflow series had energy densities located above the 99% confidence level. Energy densities for IMF1 and IMF2 in the Zhimenda streamflow series were located between the 85% and 90% confidence level. IMF3 was highly significant, whereas IMF1 and IMF2 were moderately significant. IMF4 in the Zhimenda streamflow series was located below the 85% confidence level line. The period identification of IMF4 may have been seriously masked by the trend and white noise.

Four IMF components were also extracted from the annual streamflow time series for the Tangnaihai hydrological station during 1960–2012 (Figure 4(a)). IMF1 had an approximately 4-year cycle, and IMF2 had an average period of approximately 9 years. IMF3 and IMF4 had approximately 18-year and 45-year cycles, respectively. Also shown in Figure 4(a) is the EEMD residual component, which indicates that streamflow at the Tangnaihai hydrological station had an increasing trend before 1984 and subsequently exhibited a decreasing trend. The downward trend for streamflow that began in the middle and latter half of the 1980s appears in most parts of the Yellow River basin and is generally believed to be caused by decreasing precipitation and the intensification of human activity (Miao et al. 2012).

Figure 4(b) demonstrates that IMF1 in the Tangnaihai streamflow series had energy densities located between the 85% and 90% confidence levels. Energy densities for IMF2 were located between the 90% and 95% confidence levels. IMF3 had energy densities located above the 95% confidence level, and IMF4 was located below the 85% confidence level. The period identification for IMF4 in the Tangnaihai streamflow series may have also been seriously masked by the trend and white noise.

Figure 5 shows the dominant frequencies, long-term trends, and changes in annual streamflow time series for the Changdu hydrological station. For 1960–2012, four IMFs with different periods and an overall trend were identified using EEMD (Figure 5(a)). IMF1 had an approximately 4-year cycle, and IMF2 had an average period of approximately 7 years. IMF3 and IMF4 exhibited approximately 12-year and 27-year cycles, respectively. The final residual indicated a clearly decreasing trend in the Changdu streamflow series.

Figure 5(b) shows that IMF1 in the Changdu streamflow series had energy densities located between the 85% and 90% confidence levels. Energy densities for IMF2 were located between the 95% and 99% confidence levels. IMF3 had energy densities located below the 85% confidence level line. Period identification for IMF3 in the Changdu streamflow series may have been seriously masked by the trend and white noise. IMF4 in the Changdu streamflow series was located above the 99% confidence level line.

The relationships between large-scale atmospheric circulation systems and the regional hydroclimatic situation are complicated, indicating that the atmospheric circulation–streamflow teleconnections are weak. In comparison with nonlinear statistical methods, linear correlation can be a simple and efficient method for evaluating atmospheric circulation–streamflow teleconnections (Chiew & McMahon 2002). Therefore, linear correlation was adopted to identify the main atmospheric circulation factors of streamflow variability in the three TRHR subregions.

Table 2 lists the correlation coefficients between the normalized indices of streamflow in the TRHR and large-scale atmospheric circulation during 1960–2012. The AO, SOI, EAM, and SAM indices were positively correlated with the normalized indices of streamflow at the Zhimenda hydrological station located in the SYR. The NAO and TPM indices have negative relationships with the normalized indices of streamflow at the Zhimenda hydrological station. The largest absolute value of the correlation coefficient was discovered between the normalized indices of streamflow and the TPM. Together, these results suggest that the local circulation of the TPM is the main factor driving streamflow changes in the SYR.

The AO, NAO, SOI, SAM, and TPM indices are positively correlated with the normalized indices of streamflow at the Tangnaihai hydrological station located in the SYeR. The EAM index has almost no relationship with the normalized indices of streamflow at the Tangnaihai hydrological station. In the SYeR, the largest absolute value among the correlation coefficients was discovered between the normalized indices of streamflow and the AO index. Together, these results indicate that AO is the main factor driving streamflow changes in this region.

In the SLR, the AO, NAO, SOI, SAM, and TPM indices are positively correlated with the normalized indices of streamflow at the Changdu hydrological station. However, the largest absolute value of the correlation coefficients was discovered between the normalized indices of streamflow and the SAM index. Therefore, the SAM is the main factor driving streamflow changes in the SLR.

To elaborate on the multiscale links between streamflow in the TRHR and large-scale atmospheric circulation systems, cross-wavelet and wavelet coherence analysis methods were applied to detect relationships between the normalized indices of streamflow and the main large-scale atmospheric circulation factors.

Figure 6 depicts the cross-wavelet and wavelet coherence power spectra for streamflow at the Zhimenda hydrological station and the TPM indices. At interdecadal timescales, the cross-wavelet power showed a 6- to 11-year band during the approximate period of 1975–1985; a coherence phase in the 7- to 14-year band during the approximate period of 1979–2000 was detected in the wavelet coherence power spectra. Phase relationships between streamflow and the TPM indices were unstable in this region. At interannual timescales, a coherence phase was detected in the 5- to 7-year band during the approximate period of 1991–2009. Phase changes that were significant at the 95% confidence level were dominated by antiphase relations. The cross-wavelet and wavelet coherence power spectra indicated that the periods of 3, 6, and 9 years in the Zhimenda streamflow series are related to the TPM cycle.

Figure 7 shows the cross-wavelet and wavelet coherence spectra for streamflow at the Tangnaihai hydrological station and the AO indices. At interannual timescales, the cross-wavelet power showed a 3- to 5-year band during approximately 1968–1973 and a 2- to 5-year band during the period of approximately 1987–1992. A coherence phase in the 3- to 6-year band at approximately 1965–1975 and another coherence phase in the 3- to 5-year band during approximately 1985–1992 were detected in the wavelet coherence power spectra. At interdecadal timescales, a coherence phase was detected in the 7- to 11-year band roughly during 1969–1992. Phase changes that were significant at the 95% confidence level were dominated by in-phase relations. The cross-wavelet and wavelet coherence power spectra indicated that the periods of 3 and 7 years in the Tangnaihai streamflow series were related to the AO cycle.

The cross-wavelet and wavelet coherence spectra for streamflow at the Changdu hydrological station and the SAM indices are displayed in Figure 8. At interannual timescales, the cross-wavelet power showed a 5- to 6-year band around 1972–1978 and a 5- to 7-year band around 1996–2002. A coherence phase in the 2- to 5-year band around 1970–1975 was detected in the wavelet coherence power spectra. Phase changes indicate that an in-phase relationship exists between streamflow and the SAM indices. The cross-wavelet and wavelet coherence power spectra indicate that the periods of 4 years in the Changdu streamflow series are related to the SAM cycle.

Different trends in TRHR streamflows detected using EEMD residuals may be caused by variations in the effects of prevailing climate oscillations, watershed environmental settings, and human activities (Cuo et al. 2014).

Precipitation is the principal source of water in the rivers and is the main driver of variability in the water balance over space and time. Therefore, changes in precipitation have serious implications for streamflow. Correlation coefficients between the normalized indices of precipitation in the TRHR and large-scale atmospheric circulation were calculated to verify the relationship between streamflow and large-scale atmospheric circulation systems (Table 3).

In the SYR, the TPM and SAM play key roles, resulting in correspondingly high correlation coefficients. This is particularly relevant for the TPM, which has a negative correlation coefficient of 0.37 (p < 0.05) with precipitation. Precipitation in the SYeR had a positive correlation with the AO, with a correlation coefficient of 0.33 (p < 0.05). Precipitation in the SLR had a positive correlation with the SAM, with a correlation coefficient of 0.28 (p < 0.05). In years with a strong SAM, the amount of precipitation was high.

The relationship between precipitation and large-scale atmospheric circulation confirmed the relationship between streamflow and large-scale atmospheric circulation. Regional differences exist in terms of the effects of large-scale atmospheric circulation changes in TRHR streamflow. The TPM, AO, and SAM are the key factors affecting streamflow in the SYR, SYeR, and SLR, respectively.

The TPM exerts a sizable influence on the regional climate as a result of thermal forcing (Lin & Wu 2011). Thermal forcing over the Tibetan Plateau plays a considerable role in regulating climate patterns and variations over and around the plateau, resulting in climate change in the SYR (Duan et al. 2013). A significant 4- to 6-year period in the TPM indices is apparent in the early 1990s (Ma & Gao 2003), suggesting that a mainly antiphase relationship exists at interannual timescales between streamflow at the Zhimenda station and the TPM indices.

The AO significantly influences the moisture budget in mid- to high-latitude regions in the northern hemisphere. The AO has periods of approximately 3, 8, and 12 years (Jevrejeva et al. 2003). Wavelet coherence spectra indicate excellent agreement between the periods of largest streamflow variance in the SYeR and the periods of significant power in the AO.

The SAM is a regular, annual phenomenon that brings heavy rainfall into the Lancang River Basin. The SAM brings southwest airflow into the basin and forms the dominant airflow direction and water vapor source in the region. At interannual timescales, the SAM occurs over periods of approximately 2, 4, and 7 years (Ding et al. 2013). Wavelet coherence spectra indicate excellent agreement between the periods of largest streamflow variance and SAM periods with significant power.

Research on the multiscale variability of streamflow and its links to large-scale atmospheric circulation systems in the TRHR helps with understanding local water resource changes and assessing the effects of climate change. It provides scientific support for the urgency of implementing environmental protection measures and the sustainable development of water resources in the TRHR. Using EEMD and wavelet analysis methods, this study analyzed the multiscale variability of streamflow and links to large-scale atmospheric circulation indices during 1960–2012 in the TRHR. The following conclusions can be made.

• (1)

  Trends in the TRHR streamflow vary by region. A steady increasing trend was detected in the streamflow of the SYR, and a steady decreasing trend was detected in the streamflow of the SLR. The streamflow of the SYeR had an increasing trend in the early years of the study period and subsequently exhibited a decreasing trend. Different trends in TRHR streamflows may be caused by variations in the effects of watershed climate settings and human activities.

• (2)

  Local TPM circulation is the main factor driving streamflow changes in the SYR. Periods of 3, 6, and 9 years in SYR streamflow were related to the TPM cycle. In the SLR, streamflow was primarily influenced by the SAM. A 4-year period in the SLR streamflow was related to the SAM cycle. The AO was the main factor influencing streamflow changes in the SYeR. Periods of 3 and 7 years in the SYeR streamflow were related to the AO cycle.

• (3)

  A mainly antiphase relationship exists between SYR streamflow and the TPM indices at interannual timescales. In the SLR, an in-phase relationship was detected between streamflow and the SAM indices at interannual timescales. At both interannual and interdecadal timescales, in-phase relationships were detected between SYeR streamflow and the AO indices.

Although our findings provide insight for improving water resource management and protection, shortcomings remain. Many factors, including land use change and other human interventions, may lead to inhomogeneity in streamflow and thus increase the amount of noise in the linkages between climate systems and streamflow. Quantitative research on links between hydrologic processes, human interventions, and climatic systems in the three subregions of the TRHR is required.

The authors greatly appreciate the excellent comments of the three anonymous reviewers for improving the manuscript.

This study was supported and funded by the National Natural Science Foundation of China (grant no. 51609008); the Key Special Project of the National Key Research and Development Program (grant no. 2016YFC0402309) and by the Asian Development Bank Technical Assistance Project (grant no. TA-9374 PRC).