Baseflow estimation in typical catchments in the Yellow River Basin, China

Baseflow is a relatively stable component of streamflow and the main source of river runoff during the dry season (Hall 1968; Tallaksen 1995; Smakhtin 2001), which generally includes runoff from groundwater or other delayed parts (Brutsaert 2005; Zhang et al. 2014). Maintaining stable baseflow plays an important role in stable water supply for human activities and ecological environmental protection in the basin (Smakhtin 2001; Gao et al. 2015; Cheng et al. 2016). At the same time, baseflow plays an important role in water resource security, water resource evaluation, water resource regulation and management, erosion sediment yield simulation, and rainfall–runoff relationship simulation.

The most commonly used scheme for baseflow is the two-component setting that considers streamflow consisting of direct flow (quick surface or rapid subsurface flow) and baseflow (flow that comes from groundwater storage or other delayed sources) (Mei & Anagnostou 2015). Direct flow generally consists of surface flow (infiltration excess or saturation excess), interflow, and rapid groundwater flow, whereas baseflow is the relatively stable flow between rainfalls (Tallaksen 1995). The separation and quantification of baseflow in a catchment is important for allocating and managing water resources, developing and utilizing water resources, calibrating hydrological and climate models, and assessing ecosystem productivity (Lott & Stewart 2016). However, it is difficult to precisely separate baseflow from streamflow. River runoff can be measured precisely by hydrological stations, but baseflow is only estimated by a variety of methods. Therefore, the separation of baseflow from streamflow has been a recurring theme in hydrology for decades (Gustard et al. 1992; Chen et al. 2008; Koskelo et al. 2012).

Many studies have been conducted to estimate baseflow by numerous methods, which have been proposed for separating baseflow from total streamflow. For example, Lott & Stewart (2016) compared six analytical methods to a mass balance method and found that the conductivity mass balance method in basin areas ranged from 27 km² to 68,117 km². Zhang et al. (2017) comprehensively evaluated four widely used nontracer baseflow separation methods against tracer-based hydrograph separation for five Eastern Australian catchments, and compared all results against the tracer-based hydrograph separation. A sliding average with rain record was developed by Koskelo et al. (2012) for daily flow data. They found that the tracer-based method calculated about one to four times more baseflow than the method they proposed for the same events, but the method was empirically based and restricted to small basin scales (<50 km²).

These baseflow separation methods are categorized into different types based on calculation principles, such as graphic method, hydrologic model method, mathematical physics method, physical–chemical method and numerical simulation method (Qian et al. 2011). The graphic method is a traditional manual method, which is simple but includes uncertainties. It is difficult to apply to the calculation of runoff in a long time series because of its strong subjectivity and low efficiency through visual judgment (Duncan 2019). The hydrologic model method and mathematical physics method require many parameters and the calculation process is complex (Xiong & Guo 2005; Chen et al. 2006). Physical–chemical methods (such as isotope tracer methods) have higher accuracy, but they generally lack spatiotemporal continuity of observation and require higher costs (Dang et al. 2011). Thus, their application is restricted. Another disadvantage is the lack of statistical analysis. Although the numerical simulation method lacks strict physical meaning, it simulates manual separation to achieve baseflow separation and can quickly obtain baseflow by programming (Huang 2007). Numerical simulation methods include single-parameter digital filtering, recursive digital filtering, HYSEP (streamflow hydrograph separation), and the smoothing minimum method (Lyne & Hollick 1979; Nathan & McMahon 1990; Chapman 1999; Eckhardt 2008). They have the characteristics of being fast, efficient, and able to batch process long sequences of hydrological data, and have been widely used in practice.

The Yellow River Basin (YRB) is located in the central northern part of China, a severely water-limited region, and baseflow has a substantial influence on this region. The river in the YRB generally has long no-rainfall or low-rainfall periods, during which the streams are supplied by released flow from groundwater storage. Therefore, as an important part of streamflow, baseflow plays an extremely important role in maintaining the YRB ecosystem. The factors that affect baseflow change mainly include climate change and human activities.

Wang et al. (2008) made a comprehensive analysis of the spatial and temporal evolution and driving factors of baseflow of the Yellow River in the past 50 years and found that baseflow accounted for about 44% of the total streamflow of the Yellow River. Bai et al. (2014) used the recursive digital filtering method to separate baseflow in the Tuwei River Basin and calculated the contribution ratio of precipitation and human activities to baseflow with the cumulative slope method. Both the recent runoff and sediment loads in the YRB have decreased significantly, with abrupt changes occurring from the late 1980s to the early 1990s (Zhao et al. 2014). Gao et al. (2011) pointed out the water in the Loess Plateau accounted for 44.3% of the Yellow River streamflow, whereas the sediment from the Loess Plateau accounted for 88.2% of the river's sediment.

To effectively improve the ecological conditions of the YRB, continuous watershed management has been applied, which is known around the world for its two schemes: the integrated soil conservation project that began in the 1970s, and the Grain for Green project that began in the 1990s (Wang et al. 2017a, 2017b). The implementation of these two large-scale projects is undoubtedly one type of geoengineering that not only mitigates climate change but also is expected to alter the hydrological cycle. With the rapid development of population, economy, and society in the region, the demand for water is constantly increasing, including domestic water, agricultural water, and industrial water (Wang et al. 2012; Zuo et al. 2016). Baseflow estimation not only provides a useful guide to understanding the hydrological cycle processes but also is necessary to develop adaptation management to solve water shortage issue in the YRB. Therefore, it is important to carry out baseflow estimation for typical catchments in the YRB.

The baseflow estimation is of great significance to the study of the water cycle and ecological environment protection in the YRB. We selected the Zuli River Basin (ZLRB), the Kuye River Basin (KYRB), the Tuwei River Basin (TWRB), and the Jingle sub-basin (JLB) as typical catchments. This paper (1) optimized the selection of eight baseflow separation methods and compared baseflow estimations in typical watersheds; (2) analyzed the inter-annual and intra-annual baseflow trend; and (3) identified the possible implications of factors that influence baseflow in the catchment.

A location diagram for the four typical basins is shown in Figure 1, and information about them is given in Table 1.

The Zuli River is a first-class tributary of the upper reaches of the Yellow River. The length of the main stream is 220.3 km, and the average slope from the river source to the estuary is 19.4‰ (Zhang et al. 2018). As a result of long-term water erosion, the valley presents a loess hilly landform landscape with an interlaced ravine.

The Kuye River is a tributary of the middle reaches of the Yellow River. The KYRB is connected to the Gushanchuan River Basin in the north and the TWRB in the south, with a main stream length of 241.8 km and an average slope of 2.58‰ (Wang et al. 2014). The hydrographic net in the basin is distributed mostly in the form of branches. The upper reaches (above Shenmu) are sand and grass land areas, and the middle and lower reaches (below Shenmu) are loess hilly and gully areas.

The Tuwei River is a tributary of the middle reaches of the Yellow River. The northern part of the basin is connected with the KYRB, and the southern part is bounded by the Jialu River Basin (Sun et al. 2010). The length of the main stream is 139.6 km, with an average slope of 3.16‰. The water systems on both sides are asymmetrically developed, with dense branches on the west bank and sparse branches on the east bank.

The JLB is located in the upper reaches of the Fen River, the second largest tributary of the Yellow River (Hu et al. 2017). The length of the main stream is 83.9 km. The average slope of the main river is 6.7‰. The recharge source of runoff in the basin is mainly precipitation.

Hydrological data at all stations were obtained from the Yellow River Water Conservancy Commission and were carefully checked and controlled before being released. We adopted daily runoff data mainly from the outlet hydrological station for each basin. We divided the time range of the Zuli River, Kuye River, Tuwei River and Jingle sub-basin into two periods. Period A was 1957–1987, 1956–1989, 1956–1988, and 1956–1987. Period B was 2006–2014, 2006–2011, 2006–2011, and 2006–2014. We adopted annual runoff data for temporal trends. We also analyzed the field flood data. The time series for the Kuye River and the Tuwei River were from 1956 to 2011, and the time series for the Zuli River and the Jingle sub-basin were from 1956 to 2014. We used daily rainfall data mainly from the precipitation stations for each basin during period B.

We used a 30 m resolution digital elevation model (Figure 1) to extract the river system and delineated the watershed boundary through a hydrological analysis, which we obtained from the Geospatial Data Cloud of the Chinese Academy of Sciences (http://www.gscloud.cn/). The resolution for the two land-use types (Figure 2) was 1 km, which we obtained from the Data Center for Resources and Environmental Sciences of the Chinese Academy of Sciences (http://www.resdc.cn/).

We compared eight baseflow separation methods: four single-parameter digital filtering methods, including Lyne–Hollick (LH), Chapman (CH), Chapman–Maxwell (CM), and Boughton–Chapman (BC) filters, recursive digital filter (Eckhardt filter, ECK), and HYSEP (fixed interval, FI; sliding interval, SI; local minimum, LM). The principles and details of these methods are described below.

On the basis of prior research (Table 2) and the characteristics of runoff in the study area, we set the parameters f₁ of the single-parameter filtering methods as 0.925, 0.95, 0.95, and 0.95, in accordance with the four method sequences. The parameters of the recursive digital filtering method were set as 0.925 and 0.5.

We obtained the average annual precipitation spatial distribution (2006–2014) by interpolation of the synergistic Kriging method (Figure 3(a). The average annual precipitation of a typical basin presented obvious latitudinal and zonal distribution characteristics, in which the northwest region and part of the eastern region of the JLB (with relatively high elevation; Figure 1) showed a distribution affected by elevation. On the whole, the average annual precipitation decreased with an increase in latitude (Figure 3(a)). An additional calculation showed that the average annual rainfall of the ZLRB, the KYRB, the TWRB, and the JLB was 340.70 mm, 387.35 mm, 422.12 mm, and 577.81 mm, respectively.

We calculated the monthly rainfall in the basin using the same method (synergistic Kriging method), as shown in Figure 3(b). Rainfall temporal distribution was mainly concentrated from June to September. The ZLRB had the lowest precipitation and the JLB had the highest precipitation, which was most obvious from June to September. Combining with Figure 3(a), the precipitation of the KYRB and the TWRB showed similar monthly distributions and similar spatial distribution patterns (zonal distribution).

The annual runoff and its trend in the ZLRB, the KYRB, the TWRB, and the JLB during the period 1956–2011 is shown in Figure 4. Compared with the four basins, the annual runoff of the KYRB and the TWRB decreased significantly. The decreasing trend of annual runoff in the four basins from high to low was in the order KYRB, the TWRB, the ZLRB, and the JLB from the linear trend line. From the 5-year slip line, the annual runoff fluctuated rather significantly in the period 1956–1980, whereas the runoff decreased sharply after this period. This phenomenon was evident in the KYRB and the TWRB. This trend is indirectly reflected in the annual runoff in the YRB experiencing an obvious decrease over the past 60 years, particularly after the early 1980s and late 1990s.

Considering the basin area, it was evident from the relative size of the annual runoff in the basin that the KYRB and the TWRB were in the midrange, the annual runoff in the JLB was the most abundant, and the annual runoff in the ZLRB was the lowest. This corresponded to the average annual precipitation of the basin between 2006 and 2014 (as shown in Figure 3).

As shown in Figure 5, we calculated the annual baseflow index from the measured daily runoff data of the study area using eight methods. Based on the datasets, the daily runoff series data were two periods, period A and period B. We found that the annual BFI tended to show a relatively steady increase, which was the opposite of the trend in runoff. Comparing the four basins, the annual BFI of the JLB was the most stable, and the annual BFI of the ZLRB gradually increased. Among the results of the annual BFI calculated by the eight methods, the results of the CH, CM, and ECK methods were relatively close, around 0.5. The results of the other five methods were generally higher, and the fluctuation was relatively large, above 0.5.

We analyzed the MEAN, SE, and Cv of the annual BFI to compare the difference of the eight methods (Table 3). From the perspective of the MEAN, the BFI estimated by the LH and BC methods was relatively larger, and the results of the HYSEP method were mostly in the midrange, whereas the estimation of the CH, CM, and ECK methods was relatively small. Moreover, the MEAN of all methods in period B was greater than that in period A. From the perspective of SE and Cv, the discrete degree by HYSEP methods was the largest, and the results of the LH method were mainly in the midrange, while the degrees of the CH, CM, BC, and ECK methods were relatively small. In addition, the SE and Cv of all methods in period B were mostly lower than those in period A. Therefore, the CH, CM, BC, and ECK methods were relatively stable, and the calculated BFI was relatively reasonable.

To further understand the characteristics of baseflow separation results obtained by various separation methods, this paper selected the year 2008, for which the BFI was closest to the average BFI, as a typical year in period B, and showed baseflow process line (Figure 6). Figure 6 shows that the fluctuation degree of the baseflow process line obtained by the eight methods was significantly different. In the stage with low flow, the difference of separation results was small. In June to October, when the runoff was large and unstable, the divided results were obviously different.

The baseflow process of the LH and BC methods showed the greatest fluctuation, and even the coincidence of the baseflow process line and streamflow process line appeared in a declining flow stage. The baseflow process line divided by the FI method was not smooth, which was inconsistent with the actual situation, because the underlying surface in the basin had damping and hysteresis effects on precipitation in the confluence stage of runoff, thus determining that the process line of baseflow should have been smooth. The baseflow processes of the SI and LM methods increased and decreased steeply with changes in runoff, and the peak values were high and sharp, which also could not reflect the actual situation of baseflow. The baseflow processes obtained by the CH, CM, and ECK methods were the steadiest and conformed to the actual underlying surface, and the baseflow process lines of the three methods were relatively close. Therefore, from the perspective of smoothness and rationality of the process line of baseflow, the segmentation results of the CH, CM, and ECK methods were relatively more accurate.

The CH, CM, and ECK methods were selected to continue analysis after comprehensive consideration of their BFI, characteristics, and baseflow process lines. To verify the reliability of the three methods and to select a suitable baseflow separation method for the study area, we separated typical storm-flood events (Figure 7). Taking the ZLRB as an example, 070808 represented the beginning of the flood process on August 8, 2007. These floods were unimodal and bimodal floods selected from 2006 to 2014. The results of these three methods were similar, except for differences in the ascending stage of the baseflow. During the period of rapid increase of runoff, the baseflow calculated by the ECK method grew faster than the other two methods, and its peak value was higher than the other two methods. In the partial decreasing phase, the value of the ECK method was relatively small, which was more in line with the soil characteristics of the loess hilly region. Therefore, we selected the ECK method to analyze the baseflow estimation.

The inter-annual trends of streamflow, baseflow and BFI in the study area were generally different (Figure 8(a)). During the study period, annual runoff and baseflow mostly showed a decreasing trend. BFI maintained a trend of steady growth, and specific values of BFI are given in Table 3. In period A, the estimated baseflow in the four basins fluctuated significantly, especially in JLB. In period B, the estimated baseflow in the four basins tended to be stabilize, which confirmed that the SE and Cv of BFI in period A were greater than those in period B. The corresponding BFI decreased and the annual baseflow volatility increased during the years when the annual runoff suddenly increased, such as the ZLRB in 1959 and 1964 and the KYRB in 1959.

Given the baseflow in period A fluctuated significantly, we selected the baseflow estimation in period B to analyze the intra-annual variation trend of baseflow (Figure 8(b)). Monthly variation revealed that average streamflow and baseflow were characterized by uneven distribution in a given year, especially in the ZLRB and the KYRB, and accordingly, the monthly BFI of these two basins also fluctuated greatly. The distribution of streamflow and baseflow current in the ZLRB and the JLB showed a single peak, and the peaks of streamflow and baseflow current were in August and October respectively. The streamflow distribution of the KYRB and the TWRB showed a bimodal pattern during the year. The bimodal months of streamflow were March and August, respectively, for the KYRB and the TWRB. The peak months of the KYRB baseflow were March and September, but the distribution of the TWRB baseflow was relatively uniform within the year. When streamflow and baseflow peaks appeared, such as in the ZLRB in August and in the JLB in October, the BFI of the corresponding months was relatively small.

We analyzed the relative merits of different baseflow separation methods in typical watersheds from the annual BFI and baseflow process line. The BFI obtained by the CH, CM, and ECK methods was relatively more stable and had less fluctuation, and the baseflow process line was smoother, which was consistent with the damping and hysteresis effect of the underlying surface on the precipitation in the confluence stage of runoff in the typical basin in the YRB. Different baseflow separation methods had different segmentation and calculation principles, which determined the effectiveness and suitability of the method for baseflow separation.

The segmentation results of the HYSEP method were not smooth, whereas the digital filtering method had a better separation effect, which was consistent with the underlying surface characteristics of typical basins, and in keeping with the conclusions of Dang et al. (2011) and Zhou et al. (2017). The separation principle of the HYSEP method found the minimum runoff volume to separate runoff during a period. In the period of drastic runoff change, less baseflow was included, whereas in the period of the steady runoff process, more baseflow was included, which resulted in relatively large deviations and fluctuations. The annual BFI calculated by the LH and BC methods in the single-parameter digital filtering methods was relatively large, and the BFI had a large inter-annual fluctuation. The annual BFI and baseflow process line showed that the CH, CM, and ECK methods were more suitable baseflow separation methods in the YRB, which was consistent with the conclusions of Lei et al. (2011); Wang et al. (2017a, 2017b); and Shao et al. (2020).

The CH, CM, and ECK methods were more suitable than other separation methods for typical basins in the YRB (Figures 5 and 6). Compared with period A in the ZLRB, the annual baseflow in period B accounted for a larger proportion of annual runoff, and the annual baseflow was more stable and steady. Meanwhile, the baseflow estimations for the KYRB, the TWRB, and the JLB were similar to these results. According to the comprehensive analysis of the separation of daily runoff and typical flood events, we selected the ECK method for the baseflow estimation analysis.

From the four basins comparative analysis, the annual BFI in period A was between 0.354 and 0.492, and the annual BFI in period B was between 0.443 and 0.502. The four catchments by annual BFI from high to low were the TWRB, the JLB, the KYRB, and the ZLRB. The SE and Cv of the ZLRB and the KYRB were relatively large in the four basins, indicating that their annual BFI was unstable and fluctuated significantly between these years. The annual BFI of the TWRB and the JLB was relatively stable.

Dou & Huang (2010) found that the proportion of baseflow in the six basins in the loess area obtained by the filtering method was 37–64%, and the baseflow has shown a gradually decreasing trend in recent years. Lei et al. (2011) studied the applicability of a total of eight automatic baseflow separation methods in the KYRB, including the smooth minimum method, the HYSEP method, and the digital filtering method, and calculated that the average annual BFI was 0.387. They also found that the baseflow in the KYRB has shown a decreasing trend in recent years. Wang et al. (2017a, 2017b) used four segmentation methods, namely, the BFI method, the straight-line separation method, the smooth minimum method, and the digital filtering method, to conduct baseflow segmentation in the Weihe River Basin, and the calculated average annual BFI was 0.477.

Baseflow is affected by multiple factors, such as catchment area, geology, soil, vegetation, precipitation, and land-use type (Bai et al. 2015; Shao et al. 2020). These influencing factors are interrelated yet may differ across basins with anthropogenic activities. According to the flow process line in 2008, the flow of the TWRB in the dry season was higher than that in the other three rivers, and the flow presented a horizontal lines phenomenon in March and April. Two medium reservoirs have been built on the trunk stream in the TWRB since 2000, in Yao town and Caitu ditch (Sun et al. 2010). A large amount of runoff in the flood season was dammed for agricultural irrigation and industrial water. This has not only resulted in the absence of the spring flood phenomenon like that in the KYRB, but has also made the dry season flow even higher.

The inter-annual baseflow variation in the four basins mostly showed a decreasing trend. In particular, the KYRB presented the most significant decrease (Figure 8(a)). The average annual baseflow during period B decreased by 77.25% compared with that in period A, which may be related to the effect of soil and water conservation measures and the impact of large-scale coal mining in the KYRB (Bai et al. 2015). In the YRB, soil and water conservation measures in typical basins had a significant influence on annual runoff (Mu et al. 2007). Furthermore, it had a great influence on baseflow. The change of the underlying surface affected the hydrological processes and then gradually affected baseflow. The study area land-use types from 1980 and 2005 are shown in Figure 2, and the changes in various land-use types are calculated in Table 4. The farmland in the four basins decreased to varying degrees, whereas the forest area increased, which showed the significant effects of the Grain for Green program in China. The area of water largely remained unchanged. Most of the unutilized land showed a decreasing trend, whereas increased urban and rural areas reflected the expansion of human activities. These activities, in turn, increased the consumption of water resources in the basin and reduced annual runoff and annual baseflow (Gao et al. 2011; Zhang et al. 2019).

The intra-annual baseflow distribution was as uneven as the intra-annual runoff distribution (Figure 8(b)). The maximum monthly baseflow was 0.522 × 10⁷ m³ for have been ZLRB in August, 1.04 × 10⁷ m³ for have been KYRB in March, 1.02 × 10⁷ m³ for have been TWRB in March, and 2.19 × 10⁷ m³ for have been JLB in October. Dang et al. (2011) used the single-parameter digital filtering method and recursive digital filtering method to study the baseflow separation of the upstream section of the Hei River and found that the intra-annual baseflow change of the upstream section of the Hei River increased first and then decreased, with the peak in August.

According to statistics of annual average water resources in China's first-level regions, the national average BFI was 0.26, the northwest rivers' BFI was 0.57, and the Yellow River's BFI was 0.43. The estimated BFI of typical basins in the YRB was between 0.354 and 0.502, which was broadly in line with the Yellow River's BFI. The BFI in period B was between 0.443 and 0502 and was a little higher than 0.43, which may have been due to the fact that typical watersheds had a mass of the loess hilly gully. The fractures were relatively developed and easily conducted water, which was conducive to the infiltration of precipitation and the replenishment of baseflow (Sun et al. 2010; Hu et al. 2017). The higher BFI in the study area indicated that the baseflow contributed more to river runoff and that the baseflow was an important part of water resources in the YRB (Lei et al. 2011; Bai et al. 2015).

Estimating baseflow can provide an important reference for the sustainability of water resources in the YRB. This study investigated the variation in baseflow across four typical catchments (the ZLRB, the KYRB, the TWRB, and the JLB). First, we adopted a unitary linear trend analysis method and a moving average method to analyze the trends of annual runoff. The annual runoff showed a downward trend and the annual runoff experienced an obvious decrease in the early 1980s and late 1990s. Then, we applied four single-parameter digital filter methods, a recursive digital filter method, and the HYSEP method to separate the baseflow in the four typical watersheds. The separation results, including BFI, BFI features, and baseflow process lines, showed that the CH, CM, and ECK methods were more suitable for the YRB than other methods. Moreover, the ECK method was the optimal baseflow separation method. The annual BFI of the Zuli River, the Kuye River, the Tuwei River and the Jingle sub-basin in period A was 0.354, 0.408, 0.492, and 0.480, respectively, and the annual BFI in period B was 0.443, 0.463, 0.502, and 0.496, respectively. These showed the inter-annual and intra-annual variations in the baseflow and BFI. Inter-annual baseflow showed a decreasing trend, and BFI maintained a trend of steady growth. Intra-annual baseflow was characterized by uneven distribution in a given year, especially in the Zuli River and the Kuye River, and accordingly, the monthly BFI of these two basins also fluctuated significantly. As baseflow is important for the Yellow River Basin, suitable assessment of baseflow is crucial to efficient management of water resources.

The authors would like to acknowledge the financial support received from Projects of National Natural Science Foundation of China (51979250), National Natural Science Foundation of China (31700370), National key research priorities program of China (2016YFC040240203, 2019YFC1510703), Key projects of National Natural Science Foundation of China (51739009) and Key Research and Promotion Projects (technological development) in Henan Province. We would also like to thank Emma Buckingham and the reviewers for their comments on the manuscript.

The authors declare no conflict of interest.

Data cannot be made publicly available; readers should contact the corresponding author for details.