Rainfall‐runoff responses and hillslope moisture thresholds for an upland tropical catchment in Eastern Madagascar subject to long‐term slash‐and‐burn practices

Slash‐and‐burn agriculture is an important driver of tropical forest loss and typically results in a mosaic of land uses. However, there is little quantitative information about the hydrological effects of long‐term slash‐and‐burn agriculture and how such mosaics affect the rainfall‐runoff response at the catchment scale. We monitored streamflow responses at two points along a perennial stream in a 31.7 ha catchment in eastern Madagascar that was monitored previously between 1963 and 1972. Land cover in 2015 consisted of degraded grasslands, shrub and tree fallows at various stages of regrowth (64%), eucalypt stands for charcoal production (30%), and rice paddies and wetlands in the valley‐bottom (3%). For the majority (60%) of the events during the study period, the ratio between the total amount of stormflow and rainfall was <3%, suggesting that for these events runoff was generated in the valley‐bottom only. Events for which an antecedent soil moisture storage plus rainfall (ASI + P) threshold was exceeded had much higher stormflow ratios (up to 50%), indicating that a certain wetness was required for the hillslopes to contribute to stormflow. Stable isotope sampling for four small to moderate events indicated that stormflow was dominated by pre‐event water. Total stormflow and annual water yield in 2015 were higher than in the 1960s, despite much lower rainfall in 2015. We attribute these differences to changes in soil physical properties caused by the repeated burning and loss of top‐soil, which has resulted in a reduction in the depth to the impeding layer. The changed runoff‐processes (less infiltration, more saturation‐excess overland flow) thus affect local water resources.

Plot-scale runoff responses for the different land uses depend on the K sat profile of the soil (Chappell et al., 2007;Elsenbeer et al., 1999;, which depends, in turn, on the past land use. Previous land use affects the K sat profile directly in terms of the degree of previous soil degradation and compaction, and indirectly through the loss of top-soil via erosion (Chandler & Walter, 1998;Lozano-Baez et al., 2018;Patin et al., 2012;Zhang et al., 2019;Ziegler et al., 2004).
The shallower the depth to an impeding layer, the lower the storage capacity of the soil and the greater the likelihood that hillside saturation-excess overland flow (SOF) becomes an important runoff mechanism during times of high rainfall (Birch et al., 2021b;Bonell, 2005;Elsenbeer et al., 1992;. However, the runoff response at the catchment scale cannot simply be deduced from the sum of the responses of individual vegetation patches or hillslopes (Uchida et al., 2005;Uchida & Asano, 2010).
It also depends on hillslope topography and morphology (Janeau et al., 2003;Lapides et al., 2020;Ribolzi et al., 2011), the presence of roads and footpaths (Rijsdijk et al., 2007;Sandevoir et al., 2023;Ziegler et al., 2004), and the hydrological connectivity between the different patches or hillslopes and the stream (Asano & Uchida, 2018;Ribolzi et al., 2018;Robinson et al., 1995). The degree of connectivity also depends on rainfall characteristics and antecedent wetness conditions (Ries et al., 2017;Tob on & Bruijnzeel, 2021;Western et al., 1999;Xu et al., 2019). Numerous studies have reported the existence of a wetness threshold for catchment-scale runoff responses, which is conventionally interpreted to reflect the increase in hillslope-stream connectivity as the catchment wets up (e.g., Birch et al., 2021b;Gomi et al., 2008;Litt et al., 2015;Tob on & Bruijnzeel, 2021;. Although changes in K sat , dominant flow paths, and runoff responses after converting tropical forest to pasture, agricultural cropping, or extractive plantations are well-documented (Birch et al., 2021a(Birch et al., , 2021bDe Moraes et al., 2006;de Vries et al., 2022;Liu et al., 2011;Nespoulous et al., 2019;Recha et al., 2012;Scheffler et al., 2011;Toohey et al., 2018;Zimmermann et al., 2006), integrated work in tropical catchments that have experienced multiple cycles of slash and burn agriculture is rare (Bailly et al., 1974;Ribolzi et al., 2018;Sandevoir et al., 2023;van Meerveld et al., 2019;. Therefore, we examined the effects of soil-and hillslope characteristics on catchment-scale stormflow generation in an upland catchment in eastern Madagascar that had been investigated in the 1960s by French researchers (Bailly et al., 1974) and has experienced slash-and-burn agriculture ever since. In addition, parts of the catchment are in use for charcoal production by repeated coppicing and burning of Eucalyptus trees. Our previous work in the area has revealed important differences in near-surface K sat between the different land covers (Zwartendijk et al., 2017), which result in distinctly different perched water table and overland flow responses . In this follow-on study, we examine the runoff response at the catchment scale and address the following research questions: • What are the dominant sources of streamflow during rainfall events of different magnitude and intensity?
• Is there a wetness threshold above which stormflow increases markedly, and above which hillslopes are more likely to contribute to stormflow?
• Has the rainfall-runoff response intensified since the 1960s as a result of long-term slash-and-burn activity and associated soil degradation?

| STUDY SITE DESCRIPTION
The 31.7 ha Marolaona catchment is located near Andasibe in the Ankeniheny Zahamena corridor in eastern Madagascar (18.970 S,48.422 E; Figure 1). It is drained by a 920-m long perennial stream.
Land use in the catchment and the region is dominated by small-scale subsistence farming based on slash-and-burn practices (Styger et al., 2007). Typically, rain-fed rice is grown for one or two seasons, sometimes followed by root crops like cassava for one or two more seasons before the fields are abandoned until the next slash-and-burn cycle begins. Due to increased population pressure and limited available land, the time between successive cycles in the area has decreased from the traditional 8-15 years to 3-5 years (Styger et al., 2007).
The catchment's land use in the 1960s included agricultural fields, and a mixture of shrub-and tree fallows (Bailly et al., 1974). Land use was still very patchy during the 2015-2016 follow-up study ( Figure 1) and consisted mainly of tree fallows (< 15 years old, 43%), irregularly coppiced and burned eucalypt plantations for charcoal production with an understorey dominated by Imperata cylindrica and Aristida similis grasses (30%), and shrub fallows (<5 years old, 19%). The remaining 8% of the catchment was covered by small patches of firerelated grassland (2%), semi-mature forest (>20 years old, 2%), agricultural fields (rain-fed, 1%), and irrigated rice cultivation (1%) and wetlands (2%) in the flat valley bottom (Figure 1). The stream, the rice paddies and the wetlands occupied $3% of the total catchment area (and 25%, 17% and 58% of the valley-bottom area, respectively). The dominant tree species in the tree fallows were native Psiadia altissima and Harungana madagascariensis, plus the native shrub Clidemia hirta, the fern Pteridium aquilinum and the invasive shrubs Lantana camara and Rubus moluccanus cf. Styger et al., 2007 Ghimire et al., 2022).
The elevation in the catchment ranges from 930 to 1025 m a.s.l.
Depth to the underlying Precambrian metamorphic and igneous bedrock exceeds 6 m . The soils (Tropudults) have a sandy clay loam to clayey texture (up to 45% clay at the surface, increasing slightly with depth; Andriamananjara et al., 2016). K sat decreases rapidly with depth (catchment-wide median values of $500 mm h À1 at the surface, $20 mm h À1 at 10-20 cm, and $2 mm h À1 at 20-30 cm; Zwartendijk,  and differs between land-cover types.
The highest surface K sat values are associated with eucalypt stands and tree fallows (median values of $960 and $460 mm h À1 , respectively) and the lowest with degraded grasslands (median of 42 mm h À1 ; . gauges were installed at $1.2 m above the soil surface to avoid ground-splash effects (see Figure 1 for locations). Differences in recorded rainfall amounts for the two gauges were very small (0.8% difference for total amount; cf. . Since the record for the rain gauge near the catchment outlet was complete, only those data were used in this study.

| Streamflow
Suppressed (rectangular) broad-crested weirs were constructed in the early 1960s to measure streamflow at the catchment outlet and for an upstream sub-catchment (7.1 ha) (Bailly et al., 1974; see Figure 1 for locations). We repaired the weirs and replaced the upper weir by a 90 sharp-crested V-notch. Stream water levels were measured at 5min intervals. Due to sensor failure, water level data are only available sharp-crested weir (Bos, 1989;Jan et al., 2006), validated by volumetric measurements (Supporting Information section 2.2).

| Soil moisture and groundwater levels
Volumetric soil moisture contents were measured at 5, 15 and 30 cm depth in three hillslope plots: TF2 (young tree fallow, 1.58 ha), EUC (irregularly coppiced Eucalyptus robusta trees with degraded firerelated grassland, 0.05 ha) and TSF (terraced shrub fallow, 1.93 ha) . The three depths for the soil moisture sensors were based on the rapid decline in K sat with depth below the surface . Perched water tables (PWT) were measured in three fully screened wells in the TF1 (young tree fallow, 0.05 ha), EUC and TSF plots, and one well in the TF2 plot (see Figure 1 for locations). Further details are provided by . The PWT time-series for plots TF2 (lower slope position) and EUC (mid-slope) were used for statistical comparisons with the discharge time-series.

| Water sampling for stable isotope analysis
Samples of streamflow and overland flow (lined gutters of plots TF2, EUC and TSF) were taken before, during and after rainfall events. A rainfall sample was taken for every 13 mm of rain using a sequential sampler (Kennedy et al., 1979) installed next to the recording rain gauge near the catchment outlet ( Figure 1). The sampler was emptied before 10 AM on the day after an event. In January and February 2016 (wet season) daily bulk rainfall samples were taken from the rain gauge at plot TF2 (emptied daily around 8 AM). In total, we collected 39 rainfall samples (26 from the sequential rainfall sampler and 13 bulk samples), 209 streamflow samples (173 at the lower weir and 36 at the upper weir), and 31 overland flow samples (6 at TF1, 10 at TF2, 8 at TSF and 7 at EUC).
All water samples were analysed for the stable isotopes of oxygen and hydrogen using a Cavity Ring-Down Spectroscope (L2130-i (CRDS), Picarro Inc., Santa Clara, USA) at the laboratory of the Chairs of Hydrology at the University of Freiburg (Germany). The isotope ratios were expressed relative to the Vienna Standard Mean Ocean Water using the delta notation. The stated precision was ±0.16‰ for δ 18 O and ±0.6‰ for δ 2 H. All samples plotted on or near the local meteoric waterline, suggesting that the samples were not affected by isotopic fractionation ( Figure S7).

| Event definition
To determine streamflow response characteristics, rainfall events were defined as periods with more than 5 mm of rainfall, followed by a dry period of at least three hours (cf. Ghimire et al., 2014;. For inclusion in the analysis, streamflow responses at either weir had to exceed an arbitrary threshold value of 0.02 mm h À1 . There were 25 rainfall events during the first period (15 February-13 May 2015; total rainfall: 537 mm). For two events, stream responses at the catchment outlet did not meet the criterion for minimum flow increase. Also, there were no reliable stormflow data for the upper weir for ten events during the first period due to blockage of the air vent. Hence, complete streamflow data during this period

| Rainfall characteristics and antecedent wetness conditions
For each rainfall event we computed the rainfall amount (P), duration, average intensity, the maximum and median hourly and 5-min rainfall intensities, and the time of the centroid of the event. As measures of antecedent wetness conditions, we calculated the total rainfall during the three (AP 3 ) and seven days (AP 7 ) prior to each event, stream discharge at the start of the event (Q 0 ), and an antecedent soil moisture index (ASI) for the top 30 cm of the soil at the Eucalyptus plot (ASI EUC ) and the tree fallow plot (ASI TF2 ). The ASI represents the total storage in the upper 30 cm of the soil and is calculated as the sum product of the volumetric moisture contents at the 0-10 cm, 10-22.5 cm and 22.5-30 cm depth intervals; (cf. .

| Streamflow response characteristics
For each rainfall event that produced a streamflow response, the end of the response was determined using the 'constant-k method' (Blume et al., 2007). The total stormflow amount (Q s ) was computed by subtracting a constant baseflow rate (Q 0 ) from the streamflow during the response. The stormflow ratio is the ratio between the total stormflow amount and event rainfall (Q s /P). The stormflow ratio is equivalent to the so-called Minimum Contributing Area (MCA), that is, the fraction of the catchment that would yield the measured stormflow if it contributed 100% of the rainfall it received (Dickinson & Whiteley, 1970). The total stormflow and stormflow ratio were calculated for individual events and for the two observations periods. We similarly calculated the total streamflow amount (Q) and the runoff ratio (Q/P) for the considered individual events and the entire observation period.
We, furthermore, determined for each event the maximum flow rate ( peak-flow; Q p ), and the following timing-related response characteristics: response lag time (time between the start of rainfall and the first increase in stream water level), lag to peak (time between start of rainfall and peak flow), stormflow duration (the time between the start and end of stormflow) and centroid lag time (the time between the centroids of rainfall and streamflow).

| Hydrograph separation
Only four of the sampled rainfall events had a sufficient number of stream water samples, a noticeable streamflow response, and contrasting δ 18 O values for rainfall and baseflow. For these events we applied a two-component hydrograph separation to determine the event-water (Q e ) and pre-event water (Q pe ) contributions to streamflow (Klaus & McDonnell, 2013;Sklash & Farvolden, 1979). The δ 18 O value of the baseflow prior to the event was used to represent the preevent stream water composition. For events for which only a single bulk rainfall sample was available, the isotopic composition of this sample was taken as the event-water composition. For events for which multiple rainwater samples were available, the event-water composition was based on the incremental weighted mean (McDonnell et al., 1990). The uncertainty in the event-water contribution to streamflow was calculated following the method of Genereux (1998).
We used linear interpolation of the event-water contribution (i. e., the percentage of event water in streamflow) between the individual isotope sampling times to obtain a continuous time series of event water and pre-event water von Freyberg et al., 2018) and to determine the average event-water contribution (i. e., the average percentage of event water in streamflow; Q e/ Q). We furthermore, determined the percentage of rainfall that became streamflow (Q e /P; the event-water fraction of rainfall; von Freyberg et al., 2018).

| Statistical analyses and determination of response threshold
The Kruskal-Wallis with Tukey's HSD post-hoc test was used to determine whether the median values of the response characteristics differed significantly between the two measurement locations (lower vs. upper weir) and between the two periods (Period 1 versus Period 2). Spearman rank correlation (r s ) was used to analyse the relations between event characteristics (rainfall amount and intensity, antecedent wetness) and streamflow response characteristics. Partial linear regression (Jekel & Venter, 2019) was used to determine the presence of any thresholds in these relations. Because streamflow response characteristics are influenced by event size and antecedent conditions, stepwise regression was applied to assess the influence of event characteristics on Q s , Q s /P and Q p . All statistical analyses were performed using Python 3.7, using an arbitrary level of statistical significance of 0.05.

| Determination of annual streamflow totals
For the period with missing streamflow data (14 May-11 December 2015; 212 days), total streamflow at the lower weir (i.e., the main catchment outlet) was estimated using the HBV model (Lindström et al., 1997;Seibert & Bergström, 2022), a frequently used buckettype hydrological model. The model was calibrated 50 times using a genetic algorithm by optimizing the non-parametric efficiency (Pool et al., 2018). The ensemble mean streamflow from the calibrated parameter sets was used to estimate the missing streamflow (see Supporting Information section 2.3 for details). These simulations were only used to estimate streamflow for the period with missing data to allow estimation of annual streamflow and stormflow totals for comparison with the data from the 1960s. The modelled data were not used for any other (statistical) analyses. Total rainfall for the period with missing streamflow data was 398 mm (22 rain  and February 2016 much drier (126 mm). The available streamflow data did not allow computation of annual totals for the local hydrological year (July-June; Bailly et al., 1974).

| Total streamflow
Total streamflow (Q) as measured at the catchment outlet was 372 mm during Period 1 (69% of P) and 116 mm (25% of P) for Period 2. The corresponding stormflow totals (Q s ) were 41 mm (8% of P) and 16 mm (3% of P), respectively (Table 1). Total streamflow and stormflow at the upper weir during Period 2 amounted to 100 mm (21% of P) and 27 mm (6% of P), respectively.

| Event response characteristics
The descriptive statistics of the rainfall and stormflow events are presented in Tables 2 and 3, respectively. Peak flows occurred earlier at the upper weir compared with the lower weir (Table 3 and Figure 2): T A B L E 1 Rainfall and streamflow totals at the lower weir (entire catchment of 31.7 ha) for different periods. the median lag to peak time at the upper weir was 70 min shorter than that at the lower weir. Soil moisture contents in tree fallow plot TF2 also tended to peak earlier than streamflow at the lower weir, the difference being larger (although not statistically significant) for soil moisture measured at greater depth ( Figure 2). For 12 out of 23 events, perched groundwater levels at TF2 also peaked earlier than streamflow. However, perched groundwater levels in the Eucalyptus plot, peaked later than the streamflow (by a median value of 327 min for the eight events for which there was a response in both streamflow and groundwater levels; Figure 2).
The stormflow response lag time (i.e., the time between the start of rainfall and the start of stormflow) at the lower weir was marginally correlated with AP 7 (r s = 0.29, p = 0.05) and somewhat better with rainfall event duration (r s = 0.35, p = 0.01) (Table S4) Table S5).
The median stormflow amount (Q s ) was higher for the entire catchment (1.1 mm) than for the upstream sub-catchment (0.4 mm), although the corresponding median stormflow ratios (Q s /P) were similar (Table 3). Both event total stormflow amount (Q s ) and the stormflow ratio (Q s /P) at the main outlet were strongly correlated with the antecedent soil moisture plus precipitation index (ASI TF2 + P) (r s = 0.80 and 0.84, respectively; Figure 3a, Table S4). They were also highly correlated with the maximum perched groundwater level in the tree fallow plot (PWT TF2 ; r s = 0.90 and 0.82 for Q s and Q s /P, respectively) (Figure 4a and Table S4). Total stormflow amount was also correlated with event total rainfall amount (r s = 0.78), event maximum hourly rainfall intensity (r s = 0.55), and event duration (r s = 0.51) (Table S4). Although the event total stormflow amount at the upper T A B L E 2 Summary of the rainfall event characteristics (n = 50). Maximum hourly intensity per event (mm h À1 ) 3.7 6.4 10 33 Median hourly intensity per event (mm h À1 ) 0.7 1.3 3.0 17 Note: ASI is the total amount of water stored in the top 30 cm of the soil at the start of the event.
T A B L E 3 Summary of runoff event characteristics, as measured at the catchment outlet (lower weir; n = 27) and at the upper weir (n = 34). Note: For 23 and 6 events the streamflow response was insufficient at the lower and upper weir, respectively. These events were not included.

ZWARTENDIJK ET AL.
F I G U R E 2 Box plots showing the time between peak streamflow at the lower weir (t = 0), and the time of the centroid of the rainfall, peak flow at the upper weir, peak soil moisture (SMC) at 5, 15 and 30 cm below the surface in tree fallow plot TF2, and peak perched groundwater levels (PWT) in the tree fallow TF2 and Eucalyptus (EUC) plots. Negative values indicate a peak response that is earlier than the peak flow at the lower weir. n = number of events. Note that outliers < À360 min (SMC) and >360 min (PWT) are not shown for visual clarity. Responses that share a similar letter after the name do not have a significantly different median value ( p > 0.05).
(a) (b) (c) (d) F I G U R E 3 Scatterplots of the relation between the total amount of stormflow per event measured at the catchment outlet (a, c) and at the upper weir (b, d) and the antecedent soil moisture plus event rainfall index with soil moisture measured in the tree fallow TF2 (ASI TF2 + P) (a, b) or the Eucalyptus (EUC) plot (ASI EUC + P), (c, d). The colour of the symbols reflects the maximum hourly rainfall intensity. The red line shows the best fit based on partial regression analysis. No line means that no significant relation was found.
weir was also correlated with the maximum perched groundwater level in the tree fallow plot (r s = 0.77), neither the total stormflow amount nor the stormflow ratio were correlated with ASI TF2 + P (r s = 0.28, p = 0.10 for Q s ; and r s = 0.32, p = 0.06 for Q s /P; Table S5, and Figure 3b). Event total stormflow at the upper weir was correlated most strongly with rainfall amount (r s = 0.82) and, to a lesser extent, the maximum hourly rainfall intensity (r s = 0.69), but not with rainfall event duration (r s = 0.25, p = 0.16, Table S5).
Peak flows were generally higher at the lower weir than at the upper weir (upper quartiles of 1.1 and 0.4 mm h À1 , respectively) ( Table 3). The peak flow rate was highly correlated with the total stormflow amount (r s = 0.93 and 0.77 for the lower and upper weir, respectively; Tables S4 and S5). Not surprisingly, the peak flow rates (Q p ) were also correlated with the antecedent soil moisture plus precipitation index (r s = 0.88 for the lower weir and r s = 0.52 for the upper weir). The peak flow rate was furthermore correlated with the maximum perched groundwater level in the tree fallow plot (r s = 0.84, and r s = 0.75 for the lower and upper weir, respectively), and with the maximum hourly rainfall intensity (r s = 0.44, and r s = 0.72 for the lower and upper weir, respectively).
The stepwise linear regression confirmed the importance of antecedent wetness (either represented by the ASI or pre-storm baseflow Q 0 ), and event size (P) as dominant predictors of total stormflow amount Q s . Conversely, rainfall intensity had only a small influence on the stormflow amount and peak flow rate (Tables S4 and S5).

| Thresholds for stormflow
Clear threshold relationships were identified between the antecedent soil moisture plus precipitation index (ASI + P) and event stormflow amount (Q s ) at either weir. Stormflow at the lower weir has an ASI TF2 + P-threshold value of 150 ± 6 mm (standard error) and 99 ± 3 mm for ASI EUC + P (Figure 3a, c). The ASI EUC + P-index also showed a clear threshold for stormflow at the upper weir (111 ± 1 mm; Figure 3). This was not the case for the ASI TF2 + P: the There was a similarly clear threshold in the relation between the maximum perched groundwater levels in the tree fallow plot TF2 and total event stormflow at either weir: PWT TF2 = 10.8 ± 0.1 cm at the lower weir ( Figure 4a) and 9.4 ± 0.2 cm at the upper weir (Figure 4b).
Thresholds for perched groundwater levels in the EUC plot were not significant (lower weir: r s = 0.08; upper weir: partial regression, p = 1.0; Figure 4c, d, Supporting Information section 5).

| Event water contributions to streamflow
The four rainfall events for which a two-component hydrograph separation could be applied (Table 4) covered a fairly representative range of events (cf. Table 2) in terms of rainfall amount (5-37 mm), maximum 5-min and hourly rainfall intensities (up to 120 and 33 mm h À1 , respectively), and antecedent wetness conditions (AP 3 : 0.2-53 mm; ASI TF2 : 90-144 mm). However, the stormflow ratios for the four events were small (2%-4% at the lower weir, 1%-4% at the upper weir) and the corresponding equivalent minimum contributing areas (MCA) were similar to the fraction of the catchment occupied by the valley-bottom wetlands and paddy fields (3%).
Event-water contributions for these four events ranged from 2% to 16% at the lower weir, and from 6% to 32% at the upper weir (Table 4). The maximum event-water contribution (66 ± 19%) at the lower weir was derived for the largest rainfall event (no. 48, P = 37 mm), which also had the highest 5-min rainfall intensity and the driest antecedent condition amongst the four events. The event-water contribution as a percentage of rainfall was low (0.5%) ( Table 4). The largest event-water contribution as percentage of rainfall (1%, lower weir) was found for event no. 24 (P = 34 mm), which had both the highest (ASI TF2 + P)-value (139 mm) and the highest maximum hourly rainfall intensity (33 mm h À1 ; Table 4). Figures 5 and 6 maximum hourly rainfall intensities (33 and 28 mm h À1 , respectively) and total rainfall amounts (33 and 37 mm, respectively). Total stormflow for event no. 24 was 0.6 mm (MCA: 2%) at the lower weir and 0.4 mm (MCA: 1%) at the upper weir (Figure 5a, b). Peak streamflow was higher at the upper weir (0.6 mm h À1 ) than at the lower weir (0.3 mm h À1 ). The event-water contributions during peak streamflow were 23 ± 2% and 19 ± 2%, respectively (Figure 5a, b). About 31 mm (91%) of the rain during this event fell at intensities exceeding the   Table 4) and included a burst of very high rainfall intensity (maximum 5-min intensity of 120 mm h À1 ). Peak streamflow in response to this burst was faster and higher at the upper weir (0.7 mm h À1 ) than at the lower weir (0.2 mm h À1 ). Event water contributed up 30 ± 8% of streamflow at the lower weir (no data available for the upper weir; Figure 6a, b). Rainfall subsequently continued at a lower intensity (<8 mm h À1 ), leading to a rapid recession at the upper weir ( Figure 6b) and eventually a broad second peak (0.3 mm h À1 ) at the lower weir (Figure 6a). Event water made up 10 ± 3% of this sec-

| Threshold relationships
Numerous studies in the humid temperate and tropical zones have shown that stormflow volumes and runoff coefficients increase much faster after a certain rainfall event size and/or antecedent wetness threshold is exceeded (Asano & Uchida, 2018;Detty & McGuire, 2010;Litt et al., 2015;Penna et al., 2015;Tob on & Bruijnzeel, 2021;. A few studies have shown that stormflow volumes and runoff coefficients increase with event size and rainfall intensity (Blume et al., 2007;Norbiato et al., 2009;. Thresholds for increased stormflow occurrence are usually interpreted as the onset of hillslope contributions to streamflow (Detty & McGuire, 2010;McGuire & McDonnell, 2010;Penna et al., 2015). For the Marolaona catchment, stormflow volumes seem to be primarily related to the sum of antecedent moisture in the top 30 cm of the soil and event rainfall (ASI + P), and only to a small degree to rainfall intensity (Figure 5a). This suggests that a certain soil water storage capacity needs to be filled before hillslopes contribute measurably to stormflow (cf. Penna et al., 2015;Saffarpour et al., 2016;Tob on & Bruijnzeel, 2021).
The ASI TF2 + P threshold for increased stormflow at the upper weir at Marolaona was somewhat higher during the first measurement period (February-May 2015; 147 mm) than for the (drier) second period (December 2015-March 2016, 137 mm; Table S8). Since the first period was wetter (Table 1), one would have expected the associated threshold to be lower than that for the drier, second period. This counterintuitive result may be explained by the fact that stormflow amounts at the upper weir were affected by the burning and coppicing of the eucalypt stands in the upper sub-catchment in December 2015 (i.e., during the second period). Eucalypts made up $38% of the land cover in the upper catchment, with tree fallows roughly covering the remainder (60.5%) (Figure 1). Although the water use of eucalypts under normal conditions would exceed that of the local tree fallows (Ghimire et al., 2018(Ghimire et al., , 2022Sikka et al., 2003), coppicing caused a temporary drop in soil water uptake (transpiration), leading to temporarily increased soil moisture contents, a correspondingly reduced soil water storage capacity, and hence a lower threshold value for runoff generation. Further, it cannot be excluded that burning the eucalypt fields caused surface soils to become temporarily more hydrophobic (Doerr et al., 2000) and produce infiltration-excess overland flow during times of high rainfall intensity (e.g., in January and February 2016; . Because no major changes occurred in the tree fallows of the upper sub-catchment, the observed dual relationship between stormflow and antecedent wetness in the tree fallows (Figure 3b) must be regarded an artifact of the drastic changes that took place in the eucalypt stands.
Interestingly, the (ASI + P)-thresholds associated with increased stormflow at the main catchment outlet are slightly higher than reported by  for overland flow at the plot scale: that is,150 mm vs. 137 mm for ASI TF2 + P and 98 vs.
87 mm for ASI EUC + P, respectively. Such differences across scales (cf. Spence, 2010) may reflect differences in slope gradients and/or interactions with topography (Caviedes-Voullième et al., 2021;Ribolzi et al., 2011), in particular the re-infiltration of overland flow as it moves downslope (Vigiak et al., 2008;Woolhiser et al., 1996). This reinfiltrated water may reach the stream as shallow subsurface stormflow (SSSF) or return flow, depending on slope morphology and changes in soil physical characteristics with depth (Birch et al., 2021a;Birch et al., 2021b;cf. McDonnell et al., 2021). Some support for the interpretation that hillslope SSSF rather than return flow/foot-slope SOF may be the dominant runoff contributing process at Marolaona during larger storms may come from the observation that overland flow was estimated to increase by only 0.05 mm and 0.01 mm for every extra mm of ASI + P in the TF2 and EUC plot, respectively . This is roughly an order of magnitude smaller than the corresponding increases in stormflow observed at the main catchment outlet (0.16 and 0.18 mm, for every extra mm of ASI + P in the TF2 and EUC plots, respectively; Figure 3a, c). Furthermore, correlations between stormflow amounts (at both weirs) and soil moisture contents and perched water table occurrence were much stronger than those with rainfall intensity (Tables S4 and S5).

| Runoff generation processes and hillslopestream connectivity at Marolaona
Stormflow ratios (Q s /P) and equivalent minimum contributing areas (MCA) were smaller than 3% for most events and roughly matched the percentage of the catchment covered by the stream channel and adjacent wetlands and rice paddies (Figure 1). Thus, for the $60% of all events, for which the antecedent moisture storage and rainfall amount were less than the ASI + P-threshold, stormflow was likely only generated in the riparian zone (cf. Birch et al., 2021a;Bruijnzeel, 1983;Detty & McGuire, 2010;Penna et al., 2015). During wetter conditions or larger events, the storage capacity of the soil above the shallow impeding horizon  was gradually filled and perched water tables developed (Figure 6), so that the hillslopes became increasingly connected to the stream via rapid lateral subsurface flow and saturated overland flow (Birkel et al., 2021;Bonell & Gilmour, 1978;Saffarpour et al., 2016;Tob on & Bruijnzeel, 2021;. overland flow only occurred during the largest storms (Birch et al., 2021a(Birch et al., , 2021b. The K sat profiles of the soils at Marolaona (Zwartendijk et al., 2017; imply that for more contrast is also reflected in the threshold-type relation between stormflow amount and the maximum perched water table level in TF2, whereas no such threshold was identified for the EUC plot (Figure 6). As such, stormflow at the lower weir may be related more closely to the runoff response of tree fallow sites (covering 43% of the catchment) than to sites with eucalypts (30%). This is also shown by the large difference in the lag-to-peak times between the tree fallow and eucalypts plots (Figure 4). Streamflow at the lower weir peaked after soil moisture (Figures 2c, 3c and 4;cf. Wilson et al., 1990), whereas the perched water tables at EUC peaked later (and were much less pronounced) than stormflow at the lower weir (Figures 2d and 3d). Perched water tables at EUC also occurred much less frequently than at TF2 (n = 8 vs. 23, respectively; Figure 4). All this suggests that subsurface flow from areas with coppiced eucalypts was not an important contributor to peak streamflow in the study area.
Hillslope-stream connectivity at Marolaona required the development of shallow perched water tables on the hillslopes, which were observed as far as 160 m from the stream (<40 m from the divide).
The largest stormflow amount measured at the catchment outlet (10 mm,  et al., 2023), than for the reforested catchment, where stormflow consisted mostly of SSSF and foot-slope SOF .
The results of the isotope-based hydrograph separations for the four examined rainfall events indicated that stormflow at Marolaona was dominated by pre-event water (Table 4), but event-water contributed substantially during and after peak flows for some of the events ( Figures 5 and 6). The four rainfall events were modest in size (P = 5-37 mm) and their stormflow ratios and peak flow rates as measured at the lower weir were less than the respective median values for all events (Tables 2-4). The sampled events, therefore, represent situations with a limited degree of hillslope-stream connectivity (cf. Ribolzi et al., 2018;Spence, 2010). In other words, event-water contributions might be higher for larger events with greater connectivity between the hillslopes and the stream and larger contributions by rapid SSSF and SOF. However, event-water contributions to stormflow can be relatively high for small events as well, for example, when the stormflow response is mainly caused by rain falling on the stream channel and surrounding saturated areas (Gburek, 1990). During larger storms, event-water contributions tend to be small initially (i.e., once hillslope-stream connectivity is established and SSSF supplies pre-event water), until subsurface flow and (saturated) overland flow gradually contribute more event-water to the stream as the catchment wets up (Birch et al., 2021a(Birch et al., , 2021bMuñoz-Villers & McDonnell, 2013;Penna et al., 2015;Saffarpour et al., 2016;. Streamflow responded faster at the upper weir than at the outlet (Table 2). We attribute this to the small size of the upper sub-catchment and the proximity of a wetland directly upstream of the upper weir ( Figure 1). The slower response downstream also highlights the dampening and delay of streamflow as it flows through the channel (Asano & Uchida, 2018;Uchida et al., 2005) and the modulating effects of riparian wetlands and rice paddies (Bullock & Acreman, 2003).

| Comparison with previous observations in the 1960s
Streamflow in the Marolaona catchment was measured for eight years in the 1960s. Similar measurements were made in a nearby relatively undisturbed rainforest catchment (101 ha) representing baseline conditions (Bailly et al., 1974). Rainfall during our one-year study period  (Figure 7b and S11). In the 1960s, the Marolaona catchment was dominated by dense shrubs and tree fallows, whereas in 2015, coppiced eucalypts covered 30% of the catchment (Figure 1). Before coppicing, the water use of coppiced eucalypts can be expected to be higher (Calder et al., 1993) than for shrubs and young trees (Ghimire et al., 2018;Ghimire et al., 2022) because of the more developed root system of the former (Sikka et al., 2003). Also, prior to the burning of the eucalypts in December 2015, soils under the eucalypts were generally drier and produced less overland flow (Figures 5d and 6d;  and therefore contributed (presumably) less to overall stormflow. As such, one would have expected stormflow ratios to be lower in 2015-2016 than in the 1960s. However, the opposite was found instead (Figure 7).
Although it is not clear how Bailly et al. (1974) exactly determined the end of a storm event (rendering a 1:1 comparison of stormflow ratios somewhat difficult), the stormflow ratio (Q s /P) was higher in 2015-2016 than for the two driest years in the 1960s, that is, 4% in 2015-2016 versus 1% in 1965-1966 and 3% in 1966-1967 ( Figures 7c and S11). The higher total water yield and stormflow ratio at Marolaona in 2015-2016 may thus reflect an increase in stormflow production due to progressive land degradation during the additional 40-50 years of slash-and-burn practices (Giertz et al., 2005;Toky & Ramakrishnan, 1981;Ziegler et al., 2004). Erosion during repeated cropping cycles likely stripped part of the permeable upper soil (cf. Zhang et al., 2019); see also the difference in K sat -profiles between degraded grassland and fallow sites in Zwartendijk et al. (2017) and . This reduces the soil water storage capacity of the upper soil and results in a more rapid development of perched water tables and associated production of SSSF and SOF cf. Giertz et al., 2005), thereby increasing the difference F I G U R E 7 Annual rainfall (a), total streamflow (b) and total stormflow (c) for the Marolaona catchment during the 1960s (red lines) and for this study (2015-2016, blue line with an assumed 10% measurement uncertainty for streamflow indicated in grey). For comparison, the data for a nearby forested reference catchment are shown as well (black lines). Data from the 1960s were taken from Annex II in Bailly et al. (1974). Additional information is presented in Supporting Information section 6.
in runoff response with the forested reference catchment (Figures 7c   and S11). This can affect local water resources, i.e. increase of downstream flooding and sedimentation (Brookhuis & Hein, 2016;Valentin et al., 2008).
However, due to the seasonality of rainfall at Marolaona these results need to be interpreted with additional caution. Rainfall in the wettest month (February) tends to vary greatly between years (122-449 mm between 1983 and2013;Metéo Madagascar, 2013). About 280 mm of rainfall were recorded in the second half of February in 2015, but only 55 mm during the same period in 2016. The choice of the starting and ending dates will affect the annual stormflow total (Table 1)

| CONCLUSIONS
We examined rainfall-runoff relationships in the 31.7 ha Marolaona catchment in upland eastern Madagascar, where slash-and-burn agriculture has been practiced for more than 70 years. For the majority of the examined events, the stormflow ratio was <3%. Valley-bottom wetlands and rice paddies likely generated most of the stormflow for these events. However, there was a distinct threshold relation between event total stormflow and antecedent moisture storage in the upper 30 cm of the soil plus precipitation (ASI + P). Stormflow ratios were much larger for events that exceeded this threshold, reflecting an increase in hillslope-stream connectivity during wetter conditions. Peak flow generally occurred after the onset of perched groundwater tables in the upper 30 cm of the soil on the hillslopes (mostly beneath tree fallows, the dominant land-cover). Stable isotope results for four events that were examined in more detail, indicated that stormflow was dominated by pre-event water (overall eventwater contribution ≤16%), but instantaneous event-water contributions were occasionally as high as 66%. Comparison of the streamflow responses at the catchment outlet and at an upstream weir (draining 22% of the total catchment area) suggested that the headwater catchment responded faster due to a direct connection with riparian wetlands. Land cover (notably fallows versus coppiced eucalypts) and historic soil erosion influenced the spatial pattern of stormflow generation. Despite the lower rainfall during the study period, annual stormflow and total streamflow (i.e., annual water yield) were higher than recorded for the driest years during the 1960s. This increase in stormflow is most likely due to a reduction in the soil water storage potential caused by the progressive loss of top-soil via erosion during continued slash-and-burn.

AUTHOR CONTRIBUTIONS
B. W. Zwartendijk