1. Introduction
Today, most hydrological systems are impacted by multiple factors, such as climate changes, construction of dams and/or irrigation channels, and land use [
1]. These have caused non-stationarity signals to be found in discharge time series [
2,
3,
4]. Traditionally, water-resources management has relied on the stationarity assumption; however, this can no longer be considered the default for water-resource management [
5].
To distinguish if the stationarity assumption is valid, it is necessary to consider both the mean and variance of a time-series. In a stationarity time-series, the mean and variance are constant, while in a non-stationarity series one or both measures vary. The major changes associated with non-stationaries are monotonic change and step change [
6] and are tested for using trend or change-point tests, respectively.
Many different approaches have been used to test for trends, such as the following: the Mann–Kendall (MK) test [
7]; the Seasonal MK test [
8]; Sen’s slope [
9]; the Moving Average and Spearman test [
10], and Empirical Mode Decomposition (EMD) [
11]. Tests such as the Pettitt test [
12], the Heuristic Segmentation algorithm (BG algorithm) [
13], Wavelet analysis, and the moving
t-test have been used to test for change points.
Hu et al. [
14] found the Modified MK was better at detecting trends compared to the MK test and Sen’s slope. Anghileri et al. [
15] used a combined approach to evaluate seasonal trends in the Alps, noting limited impacts on water resources. Sang et al. [
16] found EMD more effective for trend detection in China’s Hei River Basin. Zhang et al. [
17] observed consistent results across multiple methods for detecting trends and change points in the Xitiaoxi River Basin, China. Huang et al. [
18] used the BG algorithm to test for change points in the annual discharge series in the Wei River Basin in China, and found that human activities are the dominant driver of decreasing discharge. In these studies, various methods for monotonic trends and abrupt changes have been used and validated, but most employed a single approach. Therefore, multiple methods will be applied in this study to provide more objective tests results and enable an evaluation of the performance of different approaches.
Non-stationarity research in the Wei River Basin has mainly focused on annual averaged discharge to identify change and drivers [
19,
20,
21,
22]; predict the non-stationarity of discharge [
23], investigate impact on low flows and droughts [
24,
25], and investigate the flood risk [
26,
27]. Li et al. [
19] used EMD, while Zhang and Gu [
22] used the Mann–Whitney U test and the MK test; both studies found the that Baojixia channel, a water diversion channel constructed in 1971, was the most important driver of the non-stationarity of discharge. Wu et al. [
20] also investigated how the Baojixia channel affected the flow regime of the Wei River, and noted the discharge declined significantly due to the channel, particularly during the dry season. Xiong et al. [
21] found that temperature rise was the primary cause of discharge decline. Meng et al. [
23] used an SVM to forecast the non-stationarity of discharge series at three stations and found a modified empirical decomposition SVM outperformed other methods. Liu et al. [
24] observed that decreasing low flows were impacted by both climate and human activities, while Zou et al. [
25] noted that human activities dominated hydrological droughts short-term, with climate change more influential long-term. As for flood analysis, Liu et al., [
26] associated flood risk with catchment changes and extreme precipitation, while Yan et al. [
27] found similar reliability between stationarity and non-stationarity design methods. These studies mainly focused on finding the drivers of non-stationarity of different hydrological variables.
The distribution of the usage of surface water and groundwater varies significantly across the Wei River Basin [
28]. These authors state that the ratio of surface water usage to groundwater (non-agriculture) is approximately double in the upper reaches (6:1) compared to that in the lower reaches (3:1). Xu and Zuo. [
29] simulated the blue-water (discharge and groundwater) and green- water (used by the planet) resources in the Wei River basin, and found most of the basin experiences blue-water storage.
The seasonal aspect of non-stationarity in the WRB has generally been ignored. Anthropogenic and meteorological factors can both exhibit varying levels of impact across seasons. For example, Kousali et al. [
30] found that meteorological and agricultural factors impacted freshwater volume to different degrees depending on the season, in Iran.
The overall aim of this study is to investigate how non-stationarity affects water resources in the Wei River Basin at both the annual and seasonal time scales. To achieve this aim, the following research questions will be addressed:
What are the changing patterns in hydro-meteorological variables, and how do they differ at different time scales?
What are the drivers of annual and seasonal non-stationarity of discharge for analysis?
Are these associated with the regional climate?
4. Discussion
4.1. The Choice and Limitation of Non-Stationarity Detection Approaches
The importance of conducting change point detection and trend analysis simultaneously is clearly illustrated from the results. Four different non-stationarity detection methods, focusing on detecting trends and change points, were tested in
Section 3.1. Three of these methods (BG algorithm, Pettitt’s test and the MK test) can test for change points, while only the MK test and Sen’s slope can detect trends.
From
Table 3, when there is only a single change point, results from the three change-point methods are similar. However, only the BG algorithm is capable of testing for multiple change points, while the other two methods find only the single most-significant change point. To explore the impact of time-series length, the Huaxian discharge data were shortened by five years at either end of the time-series, and the results are shown in
Table 8. The length of record has a larger impact on the MK and Pettitt tests, where even slight changes in the record length results will result in different change points, especially for the Pettitt test, while the BG algorithm results are more stable.
For trend testing, the results did not vary between methods, therefore indicating that either method is suitable.
Additionally, according to Yue et al. [
35] and Farris et al. [
36], the magnitude of trend and size of sample impact the efficiency of non-parametric statistical tests. In this study, this would impact the results of the MK test and Kendall tau, with the latter used to determines the significance of Sen’s slope. Mallakpour and Villarini [
37] and Fukuda et al. [
38] also noted that the sensitivity of the Pettitt test and BG algorithm are also impacted by the length of the data record. In this study, five stations (Qinan, Weijiabao, Yuluoping, Zhangjiashan and Zhuangtou), with daily data from the last two, have fewer than 30 years of data length; therefore, the accuracy of all tests applied to these discharge time-series may be impacted, and there is the small potential for an incorrect null hypothesis to be accepted (type II error). For example, non-stationarity was not detected in the discharge time-series for Weijiabao, while significant non-stationarity signals were found at both upstream and downstream stations.
4.2. Annual Versus Seasonal Non-Stationarity Signals
Comparing the result between the annual and seasonal significant trends (Sen’s slope) for the mean time-series, there is not a consistent pattern between time-period and variable. For mean discharge, there is some consistency in results, with nine catchments showing deceasing trends at the annual scale and eight, six, eight and seven catchments having decreasing trends at the spring, summer, autumn and winter scale, respectively. The catchments with significant seasonal trends with mean discharge all have significant annual trends. There is more variability for mean precipitation and temperature. For mean precipitation, only three catchments exhibit significant decreasing trends at the annual scale, while six catchments have significant trends in spring (one increasing and five decreasing), zero in summer, four catchments with decreasing trends in autumn and three significant trends in winter (one increasing and two decreasing). Only one catchment, Linjiacun, has opposing significant trends across seasons, with a decreasing trend in spring but an increasing trend in winter. For mean temperature, autumn has the most significant trends, with eight followed by seven at the annual scale, six in winter, five in spring and three in summer. For annual, spring, autumn and winter, all trends were increasing, while for summer, one catchment (Zhuangtou) had a decreasing trend in mean temperature.
A similar pattern is also observed in the change points detected across the annual and seasonal time-periods. Nine catchments have change points in mean discharge at the annual scale and in winter; however, these are not the same catchments. Eight catchments have exhibited change points in spring and autumn, while six catchments have change points in summer. Five catchments (Beidao, Linjiacun, Xinyang, Lintong, Huaxian) exhibited change points across all periods, with the timing of these change points mostly similar across periods except, for autumn, whose change points mostly occur in 1985–1986. Only one catchment had an annual mean-precipitation change point (Linjiacun—monthly) which also corresponded to a mean discharge change point. No change points were found in the mean precipitation in either spring or summer, while four change points in autumn and five in winter were found; however, these did not align with change points in mean discharge. For mean temperature, the same eight catchments had change points at the annual scale and in both autumn and winter. At the annual scale, the change points in mean temperature corresponded to change points in mean discharge in the Beidao and Linjiacun catchment. In autumn only, in the Zhangjiashan catchment, the change points corresponded, while no change points corresponded in winter. Nine catchments had change points during spring, but none corresponded to discharge change points, and ten catchments in summer had change points, with six corresponding with mean discharge change points.
Results are similar for the maximum time-series. For maximum discharge, all significant trends are decreasing. Three catchments have significant trends across all periods (Beidao, Linjiacun and Xinyang); however, Qinan only has a significant trend in autumn and winter and Yangjiaping has a significant trend at the annual scale and in summer.
Comparing the change points of the maximum time-series, it is clear that maximum precipitation has not undergone the same changes as maximum discharge and temperature. No change points in maximum precipitation were found in the annual, spring or summer periods while only four in autumn and two in winter were detected, and none corresponded to change points in maximum discharge. Change points in maximum temperature were only found in two catchments (Beidao and Linjiacun) at the annual scale, and the timing was the same for summer but differed in both autumn and winter. Only in the Beidao, Linjiacun and Huaxian catchments did the change points for maximum temperature and discharge correspond. For maximum discharge, more change points were found for winter (eight) than any other period. Catchments along the mainstem of the WRB were found have more change points, with five of six catchments exhibiting change points at the annual scale and in spring, autumn and winter. In fact all bar two catchments (Yuluoping and Zhuangtou) exhibited change points during the winter.
4.3. The Non-Stationarity Driver Analysis and Water Resource Implications
The results obtained from trend tests and change point detection allow for a preliminary analysis of the non-stationarity drivers. Overall trends indicated decreasing trend in both discharge and precipitation across the entire basin, irrespective of annual or seasonal time-period, while temperature was increasing. This is not surprising considering the close relationship between discharge, precipitation and temperature (as a proxy for evapotranspiration); however, the number of significant trends in the mean spring series indicated that non-stationarity may have a greater impact on water resources during this period.
Change-point tests identified the fact that, in the early 1990s, change points were common along the mainstem of the Wei River and their timing corresponded to temperature change points. This indicates that temperature might be the driver of abrupt change. However, this corresponding timing is not consistent across all catchments and time-periods.
Figure 2 shows an example plot of the mean time series at the Xianyang station, located in the middle and lower reaches of the basin, which shows that the fluctuations in both the discharge and precipitation time-series align well with each other, demonstrating that there is a correlation between discharge and precipitation. Therefore, a decreasing trend in precipitation, even if insignificant, will have an impact on discharge. In addition, both the discharge and temperature time-series have abrupt change points near 1990, after which there is a sharp increase in temperature and decrease in discharge. This pattern is visible in the annual, and in all but the winter, seasonal data.
The construction of the Baojixia channel is clearly a driver of abrupt change, as the change point in 1970 at Linjiacun (daily) downstream of the Baojixia channel, which is not found upstream at Linjiacun (monthly), coincides with the construction time of the channel. From
Figure 3, it can be noticed that, since 1971, there has been a significant decrease in the daily data compared to the monthly data (~30% decrease on average), which indicated the impact caused by the water diversion. Additionally, most of the discharge stations downstream of the Baojixia channel also exhibit change points around 1970, with the except of Huaxian, which is the farthest-downstream station investigated. This indicates the scale of the impact the Baojixia water diversion project had on the WRB. This analysis supports the previous studies which also attributed the change point around the 1970s to the construction of the Baojixia channel [
19,
21,
22]. Zhang and Gu [
22]. However, the 1970s are not the dominate change point in the maximum time series, which means that while the Baojixia water diversion impacted mean discharge, it did not have the same impact on the extreme discharge.
Change points were also detected in the spring and winter mean discharge in the 1970s, as these are the dry seasons, but not in summer or autumn, indicating potential water supply issues. The results also suggest a spatial pattern in non-stationarity across the WRB, with all stations in the lower reaches having a change point located in 1959 in the mean summer mean, except those in the upper reaches. This spatial pattern may be caused by the distribution of climate zones within the WRB. From
Figure 1, moving downstream along the Wei River, the proportion of Cwa significantly increases, while the proportion of Dwb decreases.
5. Conclusions
The aim of this study was to investigate how seasonality impacts non-stationarity across the Wei River Basin. Prior to looking at the seasonality impact, it was first necessary to determine the best approach, which was carried out at the annual scale. Four different approaches (the MK test, Pettitt test, BG algorithm, and Sen’s slope) were used to test for abrupt change points and trends in discharge and meteorological factors. For the seasonality analysis, it was determined that only the BG algorithm should be used to test for change points and Sen’s slope to determine trends. The BG algorithm was used as it produced more stable results, irrespective of record length, and was the only approach that could detect multiple change points.
Overall, 410 data-series were tested for non-stationarity, of which 111 had both significant change points and trends and a further 56 had either significant trends or change point(s). When comparing change-point detection approaches, the results showed that if a time-series had a single change-point, all three methods (the MK test, Pettit test and BG algorithm) were capable of capturing this, and the timings were fairly consistent. However, only the BG algorithm was able to detect multiple change points, not just the most significant ones. The BG algorithm was also found to be less sensitive to record length.
The overall trends for the Wei River Basin are mostly consistent across catchments, in both the annual and seasonal time-series and for the mean and maximum values. Discharge experiences significant decreasing trends, as does precipitation, while temperature shows increasing trends. However, the number of significant trends varies across the annual and seasonal time-periods. There are most significant trends in winter and spring, the dry period in this basin, suggesting non-stationarity has water-resource availability. This is also supported by the result that show mean time-series are more impacted by non-stationarity than the maximum time-series.
The analysis indicated that the majority of catchments on the main-stem of the Wei River have been impacted by non-stationarity. The largest driver of non-stationarity was the construction of the Baojixia diversion channel, as change points in mean discharge correspond to its construction.
While both decreasing trends in precipitation and an increasing trend in temperature are contributing to the decreasing discharge trend, temperature appears more important. The timing of change points in temperature corresponded more regular to change points in discharge than did precipitation.
This study also explored the potential impact that different climate zones have on the non-stationarity signals detected. However, no spatial pattern was observed in the significant trends, and only twice was spatial variability observed in change points, once in the 1970s, corresponding to the construction of the Baojixia channel, and in 1959, in the mean summer time-series. Only the latter suggests any correlation with the proportion of theclimate zone within each catchment.