The studied headwater catchment, the Xitaizi experimental watershed (XEW), is situated at coordinates 40°32′ N, 116°37′ E, as depicted in Fig. . Spanning an area of 4.22 km2, XEW exhibits elevations ranging from 676 to 1201 m above sea level. Approximately 54 % of the area features a slope between 20 % and 40 %. The region experiences a monsoon-influenced semi-humid climate characterized by an average annual rainfall of 625 mm. The majority of this precipitation, around 80 %, occurs between June and September. The annual mean temperature in the area is 11.5 °C and is accompanied by a relative humidity of 59.1 %. Experimental and observational activities were conducted over the period from 2014 to 2023.

XEW represents a typical location in north China's earth-rocky mountainous region, where approximately 80 % of the catchment area is underlain by firmly compacted, deeply weathered granite. Soil mapping and field investigations reveal the prevalent soil types to be brown earth and cinnamon soil (according to Chinese soil taxonomy), with a depth extending to 1.5 m. The saturated hydraulic conductivity of the soil ranges from 19.5 to 175.3 mmh-1, with an average value of 45 mmh-1. The bedrock in the area is primarily composed of granite, constituting approximately 88 % of the total bedrock composition, while gneiss and dolomite are sporadically distributed. Some sections of the granite exhibit a fracture, and a layer of regolith is sandwiched between the soil layer and the bedrock layer. In terms of land cover, the catchment is predominantly covered by forest (98 %), with 54.2 % being broad-leaved, 2.3 % coniferous, and 10.5 % a mix of coniferous and broad-leaved. The remaining 33 % consists of shrubs (Tie et al., 2017).

Meteorological variables and runoff have been systematically monitored since 2013. Meteorological conditions were consistently measured using four GRWS100 automatic weather stations. These weather stations were strategically distributed quasi-uniformly along the elevation gradient as depicted in Fig. . The comprehensive data collection from these stations contributes to a thorough understanding of the meteorological dynamics in the study area over the specified time frame.

For the measurement of air temperature (Ta) and relative humidity at each automatic weather station, an HC2S3-L temperature and relative humidity probe was utilized. These probes were equipped with a radiation shield to enhance accuracy. Simultaneously, a LI-190R quantum sensor was employed to measure photosynthetically active radiation (PAR). Rainfall data were collected at 10 min intervals using six tipping-bucket rain gauges. These gauges were positioned in an open space near the automatic weather stations, and average values were adopted for analysis in this study.

Furthermore, the antecedent precipitation index (API), generally used to represent the residual effect of previous precipitation (Mosley, 1979; Iwagami et al., 2010), was calculated for all the events over 3, 6, and 12 d. The API during the antecedent t days is described as follows: 1API(t)=∑i=1tPii, where i is the day count and Pi is the daily precipitation on the ith day.

A Parshall flume was installed at the catchment outlet to measure streamflow (Fig. ). The water level in the flume was measured every 5 min with a HOBO capacitance water level logger from 2014. Streamflow was calculated using the standard Parshall flume-rating curve, and both the rainfall and streamflow measurements were averaged to hourly time steps, and in this study, the analysis is conducted at hourly time steps. Unfortunately, the observation equipment is susceptible to failures due to the complex environmental conditions and disturbances caused by wild animals and plants. Compounded by the remote location of XEW, accessing the site promptly to address malfunctions is challenging, leading to the loss of some observation data. Notably, stormflow data from 19 July to 16 August 2016 had to be excluded because the road collapsed during a heavy storm, preventing a significant amount of runoff from passing through the Parshall flume. Furthermore, streamflow data from 2018 to 2019 are unavailable, and the two bimodal events in 2016 were omitted from the hysteresis analysis due to substantial errors in streamflow observations resulting from damage to the diversion channel. The specific observation periods are detailed in Table . These limitations underscore the challenges associated with conducting observations in remote and environmentally intricate locations.

Volumetric soil water content (SWC) was measured at eight observation sites using CS616 time-domain reflectometry (TDR) probes at 10 min intervals. On Hillslope 1, five soil moisture sensors were deployed, with an additional three located adjacent to WS900. These sensors were strategically placed in the soil profiles at 80 cm depth intervals, each at a depth of 10 cm. For analysis in this study, the 10 min interval measurements were aggregated to hourly time steps, and the arithmetic mean of the total SWC across the four profiles was employed. Moreover, SWC data immediately preceding a rainfall event were integrated over the 80 cm depth to calculate an antecedent soil moisture index (ASI), as proposed by Haga et al. (2005). This index, commonly utilized in analyzing the impact of antecedent shallow soil water storage on catchment runoff response (Fu et al., 2013; Penna et al., 2011), provides valuable insights into the soil moisture conditions preceding rainfall events.

Fluctuations in groundwater level (below the ground surface, hereinafter referred to as bgs) were systematically recorded in eleven 80 mm diameter boreholes situated on three hillslopes within the catchment (refer to Fig. ). The boreholes were drilled to depths of 5–26 m in granite (weathered and fractured to varying extents) mantled by thin soils. Unscreened portions of the boreholes accounted for approximately one-third to three-fifths of the total depth (refer to Table ). To capture the groundwater level dynamics, HOBO capacitance water level loggers (Onset, USA) were deployed to record water levels in the boreholes at hourly intervals. It is noteworthy that water levels were rarely observed in boreholes W1-1, W1-2, W2-4, W2-5, and W2-6. This observation could be attributed to the boreholes potentially not being drilled deep enough to reach the groundwater, possibly due to challenges encountered during field drilling. Slug tests conducted following installation suggested that the saturated conductivity in the weathered and fractured granite was relatively high, ranging from 5.2 × 10-3 md-1 to as high as 1.16 md-1.

An index for groundwater level (IG) was computed by normalizing the groundwater levels in each borehole according to their recorded range throughout the research years, following the approach outlined by Detty and McGuire (2010). Subsequently, the arithmetic mean of values of IG across all boreholes was calculated, serving as a representative proxy for the groundwater level across the entire catchment. This approach provides a standardized measure that allows for the comparison of groundwater level variations across different boreholes within the study area.

An intensity-based automatic algorithm, as outlined by Tian et al. (2012) and Powell et al. (2007), was employed to delineate and segregate rainfall events from hourly rainfall time series data. In this algorithm, a threshold rainfall intensity of > 0.1 mmh-1 was utilized to determine the commencement and conclusion of each event, with individual storms being separated by a minimum of 6 h. Events characterized by an accumulated rainfall that exceeded 5 mm were selected for further analysis. A total of 95 distinct rainfall events, each with a cumulative rainfall of at least 5 mm, were identified and isolated from the rainfall data series spanning the years 2014 to 2023 by employing the intensity-based automatic method (refer to Table ).

Storm runoff events are identified when streamflow experiences a rapid increase and attains a peak in response to a rain impulse. Throughout the analyses presented, streamflow refers to the total discharge measured at the weir. The computer program HYSEP (Sloto and Crouse, 1996) was employed to automatically partition a streamflow hydrograph into baseflow and stormflow components. Subsequently, the automated separation outcomes underwent manual verification and adjustment, aligning with observed data and widely accepted straight-line separation principles. In the context of each event, q0 is defined as the streamflow before the onset of rainfall. This parameter characterizes the baseflow conditions preceding the hydrograph's response to a rain impulse (Zillgens et al., 2007). The separation of stormflow from baseflow allows for a more detailed examination of the runoff dynamics during distinct rainfall events.

The hydrograph served as a valuable tool for characterizing the timing, magnitude, and duration of runoff responses to rainfall. Three primary response types were identified based on the number and shape of streamflow peaks: unimodal, bimodal, and hybrid bimodal events. Schematic diagrams illustrating these three types of events are presented in Fig. .

A unimodal event has a single peak that generates during or shortly after the cessation of rain impulse (refer to Fig. a), while a bimodal event features two peaks as a response to the same rain impulse, of which the direct peak (also called the first peak) corresponds to a fast catchment response to rainfall and occurs synchronously with the rainfall or shortly after its onset. Additionally, we observed that the hydrographs of these events resembled those of unimodal events but exhibited significantly greater water yields and longer peak delay times. We termed these events “hybrid bimodal events”. Hybrid bimodal events can be distinguished from unimodal events by their extremely high streamflow volume, longer duration, and delayed response time (Fig. c). For the hydrographs of bimodal and hybrid bimodal events, refer to Fig. .

It is worth noting that a rainfall event may consist of multiple impulses, and in such cases, the hydrograph responds with multiple direct peaks (see Fig. b). The stormflows from the first peak (q1p) and delayed peak (q2p), along with the total event stormflow (qs = q1p+q2p), were calculated by summing hourly values over the identified event period. The runoff ratio (Rr), commonly used to estimate the effective contributing area during a runoff event (Buttle et al., 2004; Detty and McGuire, 2010), is calculated as the ratio of qs to gross rainfall.

The lag time, defined as the duration between peak rainfall and peak streamflow (Mosley, 1979), is a critical parameter for modeling the temporal variability in streamflow. Lag time varies significantly among different water sources (Becker, 2005; Haga et al., 2005) and has been introduced to comprehend sub-components of runoff in different response processes. In this study, two specific lag times are considered: t1p, the time lag between peak rainfall intensity and the first streamflow peak, and t2p, the time lag between peak rainfall intensity and the delayed streamflow peak, as illustrated in Fig. .

Water samples for isotope analysis (δ18O and δD) were collected from 1 July to 1 September 2021. Rainwater was automatically sampled every 2 hours using an ISCO6712 automatic water sampler (Teledyne ISCO Inc., Lincoln, Nebraska, USA) positioned near the weir. Manual bulk samples of rainfall were also collected at the same location after each event using a rainwater sampler with a 9.5 cm diameter funnel attached to a 500 mL plastic water bottle insulated with bubble foil to protect against direct sunlight and a table tennis ball placed in the funnel's mouth to minimize evaporation.

Stream water was collected every 2 hours upstream of the Parshall flume location using an automatic water sampler (Fig. ). Spring, seepage water, and groundwater were manually collected daily from boreholes using a bailer. All collected samples underwent isotopic composition analysis (δ18O and δD) using a Picarro L2140-i isotopic liquid water and water vapor analyzer (wavelength-scanned cavity ring-down spectroscopy, WS-CRDS) with a declared precision of δ18O ± 0.1 ‰ and δD ± 1 ‰.

To trace the source of the streamflow during storm events, a simple mass balance approach was employed to segregate the streamflow into two components: event water and pre-event water. These components are represented by rainfall and baseflow, respectively, based on the oxygen isotopic concentration (δ18O) of each component. The δ18O of baseflow and weighted rainwater samples served as end members, defining the ultimate isotopic composition of the stream, in accordance with the approach outlined by Padilla et al. (2014), as follows: 2Cs=xCe+(1-x)Cp,3x=Cs-CpCe-Cp⋅100[%], where Cs, Ce, and Cp refer to δ18O concentrations of stream, event, and pre-event water components, respectively. Ce is the weighted value calculated using the incremental weighted mean method (McDonnell et al., 1990) for each event. Cp is determined from the stream δ18O concentration measured immediately preceding the rainfall. Finally, x is the percentage of event water in the stream.

During the period from 2014 to 2023, a total of 95 distinct rainfall events, each with a cumulative rainfall of at least 5 mm, were identified from the rainfall data series. Among these events, 14 exhibited a bimodal response and an additional 4 displayed a hybrid bimodal process (refer to Table ).

The stormflow volume and lag times of streamflow peaks for both unimodal and bimodal events were determined and characterized. As depicted in Fig. , unimodal events generated relatively minimal runoff, with a maximum q1p of 0.25 mm. In contrast, the q1p and q2p of bimodal events exhibited a wider range, spanning from 0.03 to 0.38 mm and from 0.82 to 31.63 mm, respectively (Fig. b). The stormflow volume of bimodal events proved to be 3 to 114 times larger than that of unimodal events, primarily due to the presence of delayed peaks (Fig. a). Correspondingly, bimodal events displayed higher Rr values, ranging from 0.91 % to 31.81 %, whereas the Rr of unimodal events remained below 0.8 % (Fig. b). This discrepancy suggests an expanded effective contributing area during bimodal and hybrid bimodal events, as highlighted in previous studies (Zhang et al., 2021).

In both unimodal and bimodal events, all direct peaks were observed within a 1 h time frame. However, the delayed peak, a distinctive feature of bimodal events, manifested itself between 5 h and 9.9 d after the occurrence of the direct peak. Notably, hybrid bimodal events exhibited shorter lag times and significantly higher stormflow yield, underscoring the need for heightened attention in flood forecasting. The substantial difference in lag time strongly implies that these peaks are contributed by distinct water sources, aligning with findings from previous studies (Haga et al., 2005).

The relationships between different event types, rainfall characteristic parameters, and watershed wetness indicators were further depicted in Fig. . It is noteworthy that the soil water content (SWC) and groundwater level index (IG) presented in Fig.  represent data recorded at the end of rainfall events, considering that delayed streamflow peaks typically manifest subsequent to the cessation of rainfall events. Rainfall amount, IG, and SWC were statistically significantly different for both groups, as proven by the t test of the equality of medians at a significance level of α = 0.01. The transition from unimodal to bimodal events reveals a consistent increase in rainfall amount, IG, and SWC. Nearly all bimodal events exhibited rainfall amounts that exceed 50 mm, whereas the range for unimodal events varied from 5.2 to 66.6 mm (Fig. a). This suggests that the initiation of delayed streamflow peaks may be associated with substantial rainfall.

The IG and SWC of bimodal events, and especially hybrid bimodal events, were significantly higher (p < 0.01) than those of unimodal events. Despite partial overlap in the ranges of IG and SWC for these groups (Fig. d and e), the mean IG and SWC values for bimodal events (0.46 and 0.67) were notably greater than those for unimodal events (0.22 and 0.13), underscoring the distinctiveness of these parameters between event types. Contrastingly, peak rainfall intensity, mean rainfall intensity, and antecedent precipitation index (API) metrics (API3, API6, and API12) exhibited a widespread overlap in their variation (p > 0.05; Fig. b, c, and f–h). Consequently, while bimodal events were characterized by higher rainfall and antecedent wetness, IG and SWC emerged as more effective indicators for estimating the occurrence of bimodal events, while peak rainfall intensity, mean rainfall intensity, and API were found to be insufficient for distinguishing between bimodal and unimodal events.

Considering the interdependence of groundwater level, streamflow, and SWC on rainfall, a detailed examination of the relationship between rainfall amount and bimodal events was conducted. The analysis revealed that the occurrence of delayed peaks is contingent on both event rainfall and antecedent wetness, displaying a distinct threshold behavior (Fig. b). The combined sum of event rainfall amount (P) and antecedent soil moisture index prior to the rainfall (ASI0) serves as a reliable indicator for predicting the occurrence of delayed peaks. Figure  illustrates that bimodal events tend to manifest when P + ASI0 exceeds 200 mm (with only two bimodal events misplaced). An intriguing observation is that these misplaced bimodal events produced a very low value of qs and that these unimodal events near the threshold occurred just before the year's first bimodal response when the watershed was sufficiently humid, signaling a predisposition for bimodal events. However, once the rainfall surpassed the threshold, all bimodal episodes were randomly distributed, and no discernible relationship was observed between their stormflow volume (qs) and rainfall amount. Based on these findings, we posit that the stormflow generation process may be dominated by groundwater or SWC.

The preceding analysis indicates a correlation between different event types and groundwater levels along with SWC. Moreover, the inconsistent response time among different event types may signify distinct contributing sources to the stream channel, providing insights into the primary mechanisms behind runoff generation. Earlier or identical response timing of groundwater compared with streamflow suggested that streamflow response was driven by hillslope groundwater (Haught and Meerveld, 2011; Rinderer et al., 2016). To explore this further, six bimodal events with minimal or sporadic rainfall during the delayed peak period along with three unimodal events were selected. The response timing of groundwater, SWC, and streamflow is illustrated in Fig. . Each horizontal bar represents the onset of rain on the left end and the lag time of the peak value on the right end of the corresponding variable. It is worth noting that some groundwater levels in Fig. d, e, and g lack horizontal bars due to missing groundwater level data, while the groundwater levels in Fig. c lack horizontal bars due to no response from groundwater.

SWC reached their maximum after direct streamflow peaks but before delayed peaks. In typical bimodal events in particular, SWCs peaked much earlier than delayed streamflow peaks, suggesting that in these events, soil water did not contribute to direct peak but may have to delayed streamflow peaks. Regarding groundwater levels, some locations showed two peaks, and not all responded to the same rainfall event. Among different locations, groundwater levels peaked before or after the delayed streamflow peaks. However, for the hybrid bimodal events, the response time of groundwater levels at various locations and even the SWC tended to coincide with the delayed streamflow peak. Identical response timing or groundwater rising and peaking just before the stream suggest that the whole catchment or critical zone contributed to delayed stormflow.

Values of the Pearson correlation coefficient (rp) between peak groundwater levels, peak SWC, and delayed streamflow were calculated for 19 bimodal events. As showed in Fig. , the first two lines show the correlation coefficients between t2p and the lag time of the peak groundwater levels, and SWC, tG1p, and tG2p represent the response times of the first and second peaks of groundwater level or SWC, respectively. The last two lines show the correlation coefficients between q2p and the average and peak values of groundwater levels and SWC. The number after the slash specifies how many pairs of the variables.

Groundwater levels exhibited two peaks in some events, with the exception of W13. Correspondingly, among these events, the response time of the second peak of groundwater level has a strong correlation with t2p, with rp > 0.858. Even though W13's groundwater level only has one peak, this peak's response time was highly correlated with t2p at the 0.01 significance level (rp = 0.821). In contrast, SWC displayed one peak in all events, and its response time exhibited a weak correlation with t2p (rp = 0.450). Both groundwater levels and SWC, and particularly their peak values, demonstrated a high correlation with the delayed stormflow volume (qs). Above all, groundwater is deemed to be the primary controlling factor in delayed stormflow.

The robust correlation observed between groundwater levels at different locations and stormflow suggests that groundwater observations at a specific location can serve as a representative proxy for the overall groundwater level across the watershed. Given the relatively complete and dynamic water level observation data for W23, this borehole was selected for further analysis.

Considering the high correlation between streamflow and groundwater level as indicated in the previous analysis, we hypothesized a connection between groundwater and delayed stormflow. To elucidate this correlation between groundwater and streamflow, we fitted the relationship between the groundwater level at location W23 and the magnitude and timing of the delayed stormflow for bimodal events. The time lag of delayed peak (t2p) shows a negative exponential correlation with peak groundwater level (ln⁡(t2p) = 1.03 × PGL - 7.43; R2 = 0.84; p < 0.01, Fig. ), suggesting that a higher groundwater level corresponds to a faster response of the delayed runoff peak to rainfall. A comparable linear correlation was also fitted between t2p and groundwater level, albeit with a slightly lower R2 (R2 = 0.76).

Moreover, as shown in Fig. , qs also has a strong linear relationship with groundwater level (qs = -10 × PGL + 94.8; R2 = 0.91; p < 0.01). These results highlight the significant influence of groundwater on flood generation in the studied watershed, suggesting that incorporating groundwater level variations into flood forecasting models could enhance their accuracy.

For both fitted lines, the closely matching fitting lines for hybrid bimodal events support the hypothesis that these high, delayed streamflow responses, which may appear to be unimodal, are, in fact, bimodal. During hybrid bimodal events, the delayed peak increased rapidly and reached its peak within 1 d, practically merging with the direct peak. This led to a potentially misleading result that only one peak was generated. This occurrence was likely due to the groundwater level rising rapidly to a critical level with substantially higher hydraulic conductivity, allowing a larger portion of the hillslope to become hydraulically connected to the stream during these events within a very short time. Consequently, a substantial amount of groundwater was quickly discharged into the channel.

To gain additional insight into the control of groundwater level on delayed stormflow, the isotope compositions of different waterbodies were analyzed. Figure  summarizes the δ18O of stream, spring, and seepage water and the groundwater δ18O from all boreholes between 1 July and 1 September 2021. Rainwater exhibited a high variation in δ18O composition (ranging from -14.42 to -5.28 ‰), with a rainfall-weighted mean δ18O value of -9.197‰. In contrast, groundwater δ18O composition appeared more stable throughout the sampling period, showing little variation across various boreholes, with the mean δ18O value ranging from -9.76 ± 0.10 to -9.08 ± 0.86 ‰. This stability indicates minimal event-based mixing with rainwater. The δ18O values of spring and seepage water followed a pattern similar to that of groundwater. The average δ18O value of the stream (-9.51 ‰) closely resembled that of groundwater (-9.49 ‰). Although the stream's δ18O composition briefly deviated toward that of rainfall during a storm, it quickly reverted to its previous value, resembling groundwater. Large isotopic variation in rainfall was dampened in the stream, indicating that both baseflow and some stormflow originated from groundwater storage with a consistent isotopic ratio, which is a result of dispersion and mixing processes.

In Fig. , groundwater δ18O values were plotted against groundwater levels for each borehole and stream water δ18O values were plotted against streamflow. The variability in groundwater δ18O increased with rising groundwater levels, suggesting a stronger influence of rainwater on groundwater. Stream water's δ18O remained independent of streamflow volume and exhibited a range of variation similar to that of groundwater. Notably, the overlapping isotopic compositions, including those during stormflow, were predominantly found in regions with higher groundwater levels. This observation underscores that even during stormflow events, groundwater remains the primary source of streamflow.

The lag time of delayed peaks varies across different water sources, providing valuable insights into estimating stormflow water resources. Haga et al. (2005) conducted relevant studies in a forested unchanneled catchment, noting that events with shorter lag times (< 2 h) predominantly exhibited runoff that was composed of saturation excess overland flow near the spring area. In contrast, events with longer lag times (> 24 h) were characterized by river runoff mainly composed of saturated subsurface flow above the soil–bedrock interface. Becker (2005) synthesized lag times from various studies in different basins, observing a trend where lag times for the three main flow components differed by at least 1 order of magnitude following the pattern of overland flow < subsurface flow < baseflow. This substantial difference in lag times is likely attributed to the stochastic triggering of different flow paths by rainfall forcing in distinct events.

Lag times for the direct streamflow peaks, observed in both unimodal and bimodal events in this study, were generally within 30 min. These lag times exhibited no significant correlation with rainfall amount, rainfall intensity, or pre-event streamflow (correlation coefficients of 0.005, 0.017, and 0.012, respectively). This lack of correlation suggests that the direct streamflow peaks were nearly concurrent with rainfall. Therefore, we infer that these direct peaks were generated through either bypass flow mechanisms, such as macropores, fractures, or soil–bedrock interfaces, as interpreted in Buttle and Turcotte (1999), Onda et al. (2001), Uchida et al. (2005), and Xu et al. (2016). Alternatively, they could have been directly contributed to the channel by rainfall. This interpretation aligns with the consideration that the routing time of the river network in XEW is approximately 1 h (Zhao et al., 2019).

In contrast to the direct peaks, the time lags from the peak rainfalls to the delayed peaks were considerably longer, ranging from 5 h to 9.9 d (Fig. ). This lag time in our study aligns with findings from other studies where similar parameters were calculated (refer to Table ). The results imply that the delayed peaks observed in XEW were likely generated by subsurface flow processes, as indicated in the work of Lischeid et al. (2002).

For bimodal events in XEW, the non-linear relationship between groundwater level and streamflow results in hysteretic relationships between the two variables. Figure  shows time series for streamflow and IG as well as scatterplots comparing the two variables for the six events used in Sect. . As noted by Dunne (1978), when two runoff peaks appear in an event, there must be at least two zones in the catchment that respond to the storm and contribute to runoff. The hysteretic nature highlights the possibility of multiple hydrological compartments being active, and these compartments are not necessarily contributing significant flows simultaneously but rather sequentially during the runoff generation period (Fovet et al., 2015; Martínez-Carreras et al., 2016).

Streamflow increased quickly and peaked before groundwater level during direct peaks, resulting in an counterclockwise hysteretic loop. It can be explained that direct peaks were formed by rainfall directly falling onto the channel or a saturation zone near the channel and/or by the flow that contributed to the channel through rapid routes, as observed in other watersheds by Jackisch et al. (2016). In contrast, groundwater level peaked first during delayed peaks, indicating that the groundwater level in the watershed peaked first and subsequently released water, creating the delayed runoff peak. This behavior may be attributed to the groundwater level surpassing a threshold for generating bimodal hydrographs, leading to enhanced hydraulic connectivity between hillslopes and the channel. This, in turn, resulted in the swift release of a substantial amount of groundwater or subsurface flow (Burt and Butcher, 1985; Detty and McGuire, 2010; McGlynn and McDonnell, 2003; McGuire and McDonnell, 2010; Scaife and Band, 2017). Consequently, the groundwater level is not merely a passive feature in this watershed, where shallow groundwater may constitute the primary runoff component, but it actively controls the stormflow.

The two-component hydrograph separation was performed for four bimodal storm events using the δ18O of the bulk rainfall, a pre-event water signature (represented by the stream δ18O before the rainfall) and the monitored stream water signature during the events. These four events were chosen because of their relatively complete isotope data. It should be noted that in all four rainfall events, δ18O values in rain and stream water were notably different, which is a requirement for end-member hydrograph separation analysis. The hydrograph separation results as well as the δ18O series of rainwater and stream water are shown in Fig. .

Regarding the water source separation result, these four events can be divided into two groups, the first being Event B1 and B4, where the major stormflow process were lagging behind and considerably damp and event water contributions were higher compared to the other two events. The fraction of event water comprising the hydrograph was 25 % in Event B1, and the contribution ratio of event water in Event 4 was 11 %. Considering that the rain had already stopped, the event water component of the delayed peak should be the rainwater temporarily stored in the watershed during the rainfall process. Event B2 and especially Event B3, however, were almost entirely pre-event-water-dominated (the contributions of pre-event water were 92 % for Event 2 and 97 % for Event B3); although it was evident that some event water contributed to the stormflow during the rising and peak period of streamflow, this water may have originated from the direct rainfall or rainwater taking a rapid route to the stream channel.

The hydrograph separation results indicated that the streamflow contribution of pre-event water changed virtually in sync with streamflow following the onset of rain, almost entirely dominating the hydrograph, while event water dominated the sharp streamflow peak that responds to high-intensity storm. Early in the rainy event, the pre-event component of the hydrograph exceeded 50 %, indicating a sufficiently swift groundwater response such that considerable amounts of groundwater were released soon after the start of rain.

In addition, there was a noticeable, gradual rise in the pre-event water contribution to total stormflow as the catchment was wetting up (Fig. ). Event B1 had a rather dry antecedent condition and showed a relatively lower pre-event water percentage (about 75 %). Event 3 in the temporal sequence had an extremely high pre-event water proportion (of approximately 97 %) and occurred under highly wet antecedent conditions. In Event B4, due to a slightly reduced wetness condition compared to Event B3 preceding it, the percentage of pre-event water somewhat decreased to approximately 89 %. This pattern may be attributed to increased water flux during the wetting-up process, when the water table rose into near-surface soil layers with highly saturated hydraulic conductivity. The rate of groundwater increase slowed as a result of the higher transmissivity, and more pre-event water was mobilized and traveled rapidly to the stream via shallow flow pathways (Lundin, 1982).

Our field observations on site indicate that direct exfiltration of groundwater into the runoff is predominant, with few signs of hillslope overland flow. For example, during a heavy storm on 5 July 2021 characterized by short duration (7 h) and very high intensity (27.6 mmh-1), with a total rainfall of 65.2 mm, minimal overland flow was observed at the study site. However, after the storm on 5 July, the spring water flow from Hillslope 2 substantially increased. Moreover, at various points in the watershed, seepage flow was observed to be gushing from fractures in the stone and holes in the earth. These field observations strongly suggest the direct exfiltration of groundwater into the runoff, providing further support to the notion that groundwater significantly contributes to stormflow in the watershed.