The Mann-Kendall test is a widely utilized and globally recommended non-parametric approach by the World Meteorological Organization for trend analysis in hydrology and meteorology. Its popularity stems from the fact that it does not assume any specific distribution of samples and is insensitive to seasonality (Van Belle and Hughes, 1984; Nalley et al., 2013). In conducting this test, the null hypothesis (H₀) assumes that the test variables exhibit no trend during the specified period. Conversely, the alternative hypothesis (H₁) rejects H₀ in the presence of a monotonic trend within the test period. The statistical procedures involved in the Mann-Kendall test are outlined as follows (Mann, 1945; Da Silva et al., 2015): S = ∑ i = 1 n − 1 ∑ j = i + 1 n s g n X j − X i where the target statistics S is to assess trends in a time series, n denotes the total number of data points in the series, ranging from 1 to n; X _j and X _i represent the respective values at positions j and i within the time series (i < j); sgn (X _j −X _i ) indicates the difference between X _i and X _j : s g n X j − X i = + 1   X j − X i > 0   0   X j − X i = 0 − 1   X j − X i < 0

The solution to the series data variance is: V a r S = n n − 1 2 n + 5 − ∑ i = 1 p t i t i − 1 2 t i + 5 18 where Var(S) represents the normalized measure of the variance of S; p means the total group of nodes; t _i is the time range at the i node. When the number of samples are over 30 (n > 30), the standard normal statistical variable Z _S is: Z S = S − 1 V a r S , S > 0   0   S = 0 S + 1 V a r S , S < 0

In the bilateral trend test, if |Z _S | is greater than Z _1−α/2 within a given confidence level of α, it indicates the sample data have a significant trend in the testing period, which rejects the null hypothesis. If Z _S > 0, it means that the sample data have a significant growth trend in testing period, while if Z _S < 0, the sample data exhibit a significant downward trend during the testing time. When the |Z _S | is greater than 1.64 and 2.32, which means the variable is significant monotonicity at the 95% and 99% levels, respectively.

The SPEI, as a widely employed method for assessing drought conditions, relies on the trade-off between precipitation and potential evapotranspiration (Vicente-Serrano et al., 2010). Consequently, it effectively mirrors the consequences of global warming on drought patterns (Vicente-Serrano et al., 2010). Furthermore, the SPEI exhibits flexibility in adapting to the diverse time scales characteristic of drought events. (Bohn and Piccolo, 2018).

The SPEI possesses a mean value of 0 and a standard deviation of 1, serving as a reliable metric for analyzing drought characteristics. Based on factors such as intensity, magnitude, duration, and frequency, water surplus and deficit are categorized into seven distinct groups (Tan et al., 2015; Bohn and Piccolo, 2018). For the purpose of studying hydrological processes across various hydrologic years, these seven categories are further consolidated into three representative types (Table 1).

The CA-Markov model integrates the Cellular Automata (CA) and Markov Chain (MC) models, both being discrete dynamic models that operate in terms of time and state. The Markov Chain, as a stochastic model, characterizes the transition probabilities of a sequence of random variables across discrete time intervals (Sang et al., 2011; Du et al., 2012). On the other hand, CA performs spatial operations where each cell’s state at time t + 1 is influenced by its neighboring cells at time t (Ye and Bai, 1969; Mohamed and Worku, 2020). In landscapes encompassing multiple land-use categories, land-use changes exhibit the Markov property. Therefore, a land-use transition matrix is constructed to capture the probabilities of each land-use type transforming into any other type. This matrix serves as a valuable tool in predicting future land-use changes.

However, the Markov chain model ignores the significant influence of geomorphology on land use and lacks the ability to effectively capture the dynamics of land-use alternation. In contrast, the CA-Markov model leverages the spatial operation capabilities of CA and the transition probability analysis of the Markov chain, enabling it to forecast spatially explicit land-use variation over a specified period. This combined approach provides a more comprehensive and accurate understanding of land-use dynamics, considering both the spatial context and the probabilistic nature of land-use transitions (Mondal and Southworth, 2010).

The hybrid CA-Markov model was implemented using the IDRISI software platform (Eastman, 1999). Prior to predicting the land use in the Yanhe watershed for the year 2060, the model’s accuracy was rigorously tested by comparing real and simulated land-use layouts for the year 2010. Initially, a Markov transition matrix was constructed using land-use maps from 1990 to 2000. This matrix captured the probability of each land-use type transitioning to any other type. Additionally, ancillary images depicting highways, motorways, railways, waterbodies, slopes, and altitudes were employed as part of a multi-criteria decision-making process to generate a transition suitability image collection. This collection provided a spatial representation of the suitability of each land-use type. Subsequently, leveraging the real land-use layout of 2000, the CA-Markov model was utilized to generate a simulated land-use map.

SWAT is a semi-distributed, basin-scale hydrological model that incorporates various data sources to simulate the hydrological processes within a research watershed (Gassman et al., 2007; Kang et al., 2021). It relies on daily meteorological data, along with land-use maps and soil attribute data. The model divides the entire watershed into multiple subbasins and further subdivides them into hydrological response units, which share similar land-use categories, management practices, and soil properties (Neitsch et al., 2011). This localized SWAT model is capable of simulating hydrological conditions over extended periods, providing valuable insights into the watershed’s water balance and dynamics. S W t = S W 0 + ∑ i = 1 l   R day   − Q surf   − E a − W seep   − Q g w where SW _t represents the ultimate soil moisture content, indicating the final state of soil water after considering various inputs and outputs over a defined period; SW ₀ denotes the initial soil moisture content, serving as the starting point for the model’s calculations; t signifies the time phase, specifically measured in days, and encompasses the duration of the simulation; R _day , Q _surf and E _a are key components that represent daily rainfall, surface runoff, and evapotranspiration, respectively; W _seep represents the cumulative amount of water that percolates from the topsoil into the aeration zone on a specific day, marking the downward movement of water within the soil profile; Q _gw stands for return flow, referring to the water that recirculates from deeper soil layers or groundwater back to the soil surface.

Temperature serves as the most prominent indicator of climate change. Utilizing the mean annual temperature data from five weather stations, we conducted a Mann-Kendall test to analyze the temperature trend (Figure 3). The results revealed that the temperature Z _s statistic stood at 5.24 for the period 1970–2019, exceeding the Z statistic threshold of 2.32 for a 99% confidence interval. Consequently, it is evident that the Yanhe watershed experienced a significant increase in temperature over the past 50 years (p < 0.01). Furthermore, the rate of increase was 0.37°C per decade, and a notable abrupt change occurring in 1996. Specifically, the average annual temperature was 10.56°C during 1970–1996, whereas it rose to 11.80°C from 1997 to 2019. Notably, the peak temperature was recorded in 1998.

Over the past century (1905–2005), China experienced two distinct warm periods: the 1940s–1950s and the 1980s–1990s. Notably, 1998 marked the warmest year recorded in China since the inception of meteorological records (Tang and Ren, 2005). Consequently, the Yanhe watershed, influenced by global warming, exhibited similar warming trends and temperature fluctuations.

The Mann-Kendall test was utilized to ascertain the trends in precipitation and streamflow within the semiarid Yanhe watershed during 1970–2019 (Figure 4). The Z _s statistic for average precipitation in five stations stood at 0.73, falling within the range of zero to 1.64 (the threshold for a 95% confidence interval). Consequently, we observed a slight increased rate of 1.16 mm/yr inter-annual precipitation. Conversely, the Z _s statistic for streamflow was −3.09, indicating a negative trend. Its absolute value exceeded the Z statistic threshold of 2.32 for a 99% confidence interval. Therefore, it is evident that the streamflow has decreased significantly with a rate of 1.60 million m³/yr over the past 50 years.

Besides, the average streamflow in the period of 2005–2012 was only half of that with the similar precipitation during 1989–1994. It can be ascribed to ecological restoration, because runoff is the synthesis of precipitation and land surface condition. The restored vegetation from GFGP reduced the rainfall-runoff conversion and increases the infiltration of surface runoff during confluence processes. Moreover, the runoff reduction increased with the age of trees planted (Huang et al., 2003). Our findings reveal that ecological restoration lasting between 4 and 7 years can lead to a reduction of half the streamflow in the Yanhe watershed, except for extreme rainfall conditions.

To evaluate the precision of the hybrid CA-Markov, the Kappa spatial correlation—a highly regarded metric for accuracy assessment in IDRISI—was employed to contrast the actual and predicted land-use maps for 2010 (Figure 6). The analysis revealed a Kappa spatial correlation of 0.91 between the real and simulated maps, indicating a strong agreement. Drawing from the research conducted by Yang et al. (2021), a Kappa value exceeding 0.75 signifies a high degree of consistency and minimal discernible differences between the two land-use maps. Furthermore, Li and Zhou (2015) used the same method to simulate and predict the land use distribution patterns in the Yanhe watershed for 2010 and 2020, reaching similar conclusions as in this study. Consequently, the hybrid CA-Markov model, with its optimized parameters, exhibits excellent performance in predicting land-use patterns.

There are 27 parameters that influence runoff in the SWAT model (Ahn and Merwade, 2016). To localize the general SWAT, we utilized the SWAT-CUP calibration uncertainty program to identify six runoff-sensitive parameters in the Yanhe watershed. Subsequently, the optimal values for these sensitive parameters were determined through Latin hypercube sampling, ensuring a precise and customized model for the watershed (Table 2). Besides, statistical metrics including R ², Nash-Sutcliffe coefficient (NSE), and percent bias (PBIAS) were employed to assess the accuracy of the SWAT (Gassman et al., 2007; Arnold et al., 2012). The results revealed that during the calibration period (1973–1997), the R ², NSE, and PBIAS values were 0.73, 0.60, and 18%, respectively (Figure 7). Similarly, in the validation period (2000–2019), these metrics were 0.71, 0.51, and 24%, respectively (Figure 8). According to SWAT accuracy evaluation research, a model is considered acceptable when NES exceeds 0.50 and PBIAS is less than 25% for streamflow predictions (Moriasi et al., 2007). Even though the localized SWAT could satisfy the model accuracy, there exists a discrepancy between the simulated and the observed streamflow. The one possible reason of this discrepancy is that the SWAT model’s simulation of snow and ice melt is simple based temperature (Devia et al., 2015), overlooking the impacts of the land use, slope aspect and gradient. The other possible is that the precipitation in the Yanhe River Basin, located in a semi-arid region, exhibits significant interannual variability, and the study period spans a relatively long duration, which leads to the basis between the simulated and observed streamflow. This research focuses on the impacts of land-use change on hydrological processes during the flooding seasons, and selects consecutive typical hydrological years to drive the localized SWAT model. Therefore, the calibrated model is able to reflect the characteristics of the impact of land-use variation on hydrology during typical hydrological years.

According to the SWAT land-use classification system, the Yanhe watershed’s land use was reclassified into grassland, farmland, forest, urban/built-up land, and water body. An analysis of land-use changes from 1990 to 2010 revealed several notable trends (Table 3). Specifically, the areas of grassland, forest, and built-up land increased by 9.38%, 1.54%, and 0.40%, respectively, while farmland and water areas decreased by 11.17% and 0.16%, respectively. When comparing the land-use changes between 1990 and 2010 in greater detail, it was observed that from 1990 to 2000, the areas of grassland, forest, built-up land, and water body increased by 0.70%, 0.13%, 0.18%, and 0.01%, respectively, while farmland decreased by 1.03%. However, from 2000 to 2010, the changes were more significant, with grassland, forest, and built-up land continuing to expand, while farmland continued to decline. Overall, the most prominent land-use changes in the Yanhe watershed from 1990 to 2010 were the increasing trend in grassland, forest, and built-up land areas and the decreasing trend in farmland.

Using the calibrated IDRISI model with a land-use transition matrix derived from data between 2000 and 2010, predictions were made regarding the land-use layout anticipated for 2060 (Figure 9). When comparing the predicted land-use state of 2060 with the actual state of 2010, it is evident that the most significant change would involve the transformation of 9.21% of grassland. This grassland would primarily be converted into 5.63% of forest and 3.58% of built-up land. Notably, the expansion of forestland is expected to occur primarily in the midstream and downstream regions of the basin. Conversely, the growth of built-up areas is predicted to be concentrated in gullies with flatter terrain and more favorable urban habitat environments.

SWAT models were constructed utilizing five consecutive, representative drought years (comprising two warm-up years and five simulation years) alongside land-use maps from 2010 to 2060 within the Yanhe watershed. The findings revealed that the impact of arid climatic conditions on streamflow was primarily concentrated during flooding periods. Specifically, the mean discharge rates during these periods were 5.05 m³/s and 6.69 m³/s for the 2010 and 2060 scenarios, respectively. Furthermore, the average flood peak in the 2060 landscape scenario exhibited a notable increase of 32.49% compared to that of 2010 (Figure 10A). Overall, during these typical drought years, the Yanhe watershed exhibited low discharge levels and rapid fluctuations in flood peaks.

The significant land-use alteration from 2010 to 2060 is the grassland converted to the forest and built-up land. The transformed forest is primarily located on hilly slopes, situated far from the Yanhe River. Conversely, the converted built-up land is situated in proximity to the river, occupying a broad and flat area. Consequently, the surface runoff generated by limited precipitation on the built-up land drains into the river, leading to a rapid response of low streamflow to rainfall during typical drought years. We refer to this phenomenon, where low streamflow promptly reacts to limited precipitation in close proximity to the land surface, as the “local rainfall-runoff effect".

Utilizing the 2060 landscape as a basis, we conducted a thorough analysis of the intra-annual characteristics of precipitation, runoff, and baseflow to further elucidate the “local rainfall-runoff effect” during typical drought years (Figure 10B). Our findings revealed that rainfall was the primary determinant of surface runoff in these drought-prone years, especially during flooding periods. Specifically, total precipitation amounted to 319.75 mm, with 56.74% of this precipitation occurring exclusively during flooding periods. Notably, 64.66% of this precipitation was converted into surface runoff, serving as the principal mechanism for the rapid increase in streamflow during flooding periods. On the other hand, baseflow contributed 50.72% to streamflow, becoming the primary source of recharge for the river during non-flooding periods.

The aforementioned analysis clearly illustrates the “local rainfall-runoff effect”. This effect refers to the phenomenon where a significant portion of limited precipitation generates surface runoff on impermeable or weakly permeable surfaces. This runoff subsequently recharges nearby streams and rivers, resulting in a rapid increase in discharge. Therefore, the local landscape characteristics play a pivotal role in influencing the streamflow patterns during typical drought years.

According to the classification of SPEI, we selected five consecutive normal years (consisting of two warm-up years and five simulation years) to forecast hydrological variations within the Yanhe watershed. Our findings revealed that the influence of a normal climate on surface runoff was primarily concentrated from June to October. When comparing the average discharge based on the landscapes of 2010 and 2060, we observed a decrease from 19.42 m³/s and 12.40 m³/s. Notably, during these normal years, the flood peak with the 2060 land use scenario decreased by 36.15% compared to that with the 2010 land use (Figure 11A). Nevertheless, it is worth mentioning that the flooding discharge in normal years is approximately twice as high as that observed during typical drought years.

When comparing the land-use state in 2010 with that of 2060, we observe that the prolonged restoration of forests and grasslands in 2060 effectively limits upstream and local surface runoff generated by relatively abundant precipitation during normal years. This, in turn, reduces the runoff flows into the river. Although the expansion of built-up land (accounting for 4.49% of the entire watershed) converts a significant amount of rainfall into surface runoff, the peak flow in 2060 remains lower than that in 2010. This is attributed to the fact that the surface runoff restricted by restored vegetation outweighs the runoff generated by the built-up land. Furthermore, given the high discharge rates during normal years, the limited surface runoff generated on the relatively small percentage of built-up land has minimal impact on streamflow. Consequently, when compared to the landscape of 2010, the peak flow in 2060 is lower. We refer to this phenomenon as the “global rainfall-runoff effect".

Utilizing the land-use map of 2060, we conducted a thorough analysis of the intra-annual distribution of precipitation, runoff, and baseflow during normal years to elucidate the “global rainfall-runoff effect” (Figure 11B). Our findings revealed that precipitation, streamflow, and baseflow exhibited smooth transitions throughout the normal years, with the timing of peak flow lagging behind the period of maximum precipitation. Notably, the peak volume of precipitation occurred in July, whereas the peak discharge was delayed until early August. On average, the total precipitation amounted to 479.66 mm, with only 3.99% of this precipitation converting into surface runoff and subsequently entering the stream during these normal years.

There are two primary reasons for the observed lag between the flood peak and the maximum precipitation time. Firstly, the smooth fluctuations in rainfall distribution throughout the year maintain the discharge within a consistently high range, rendering it less susceptible to the “local rainfall-runoff effect”. Secondly, the well restored vegetation on the slope effectively intercepts surface runoff, and the homogeneous loess soil has high infiltration capacity in the normal hydrological years (Jin et al., 2020). Consequently, when rainfall is relatively abundant and the river maintains a stable discharge, the rainfall-runoff process is primarily influenced by the spatial distribution of major land-use types across the entire basin, giving rise to the “global rainfall-runoff effect”.

For a comparative analysis with typical drought years and normal years, the SWAT models were driven using data from five consecutive typical wet years. These models were based on the land-use layouts of 2010 and 2060 in the Yanhe watershed. The results indicated that climate was the primary determinant of streamflow in flooding periods of wet years. Specifically, the flood peaks were virtually identical for both the 2010 and 2060 land-use scenarios. During the typical wet years, the discharge was high, and the peak flow was predominantly concentrated in July (Figure 12A).

Previous studies indicated that vegetation (forest and grass) has a good function of soil and water conservation (Jin et al., 2020; Lyu et al., 2023). The area of vegetation in the Yanhe watershed accounts for 82.00% and 78.43% of the entire watershed in 2010 and 2060, respectively. However, it has limited capacity to restrict surface runoff during intense rainstorms or prolonged rainfall. The aggressive surface runoff resulting from extreme rainfall is scarcely intercepted and infiltrated by the forest and grassland. This was evident during the “7.26” rainstorm event in the Wudinghe River basin of the LP on 26 July 2017. Despite decades of ecological restoration efforts, where the area of grassland and forest has exceeded 50% of the watershed, the rainstorm still generated significant flood peaks and caused severe soil erosion (Tang et al., 2020; Wang et al., 2020). Consequently, in typical wet years, the aggressive surface runoff rapidly flows into the river, significantly increasing the peak discharge.

To gain a deeper understanding of the hydrological processes during typical wet years, we analyzed the intra-annual distribution of rainfall, runoff, and baseflow (Figure 12B). The results indicated that the average precipitation in these wet years was 624.01 mm, with significant variations within the year. Notably, approximately 55.38% of the total precipitation occurred primarily in July and August. Concurrently, the surface runoff showed a marked increase during the concentrated rainstorms in July, accounting for a substantial 79.11% of the total surface runoff. During this period, the flood discharge was particularly high and declined rapidly following the intense rainfall. As approximately 90% of the Yanhe watershed comprises hilly-gully regions (Wang et al., 2016), the occurrence of rainstorms frequently triggers infiltration-excess overland flow (Jin et al., 2020), and the resulting stormflow responds promptly to precipitation.