Large broad-leaved canopy of banana (Musa nana Lour.) induces dramatically high spatial–temporal variability of throughfall

Throughfall (TF) is an important water input of rainfall redistribution into floor, and its spatial–temporal variability under some species’ canopies has been documented to evaluate effect on splash erosion. However, the understanding of TF variability under large broad-leaved canopy remains insufficient. In this study, the spatial heterogeneity, temporal stability and drop size of TF were quantified using variogram fitting, normalised ranking and filter paper staining, respectively, under banana (Musa nana Lour.) canopy comprising long and wide leaves. Results indicated TF pattern showed strong spatial correlation at a range of 3–5 m. High spatial variability of TF was found, which was affected by rainfall event size and was accompanied by great canopy disturbance. TF plots revealed high time variability, which was mainly controlled by unstable banana canopy structure. TF drop size from leaf dripping points varied in 3–10 mm and showed significant differences (p, 0.05) among five kinds of leaf shapes, implying that concave and broken banana leaves were involved in the variability of TF drop size. Overall, results demonstrate the spatial–temporal variability of TF is dramatically induced by banana canopy with broad leaves, which may result in non-uniform soil water content and splash erosion under the canopy.


INTRODUCTION
Throughfall (TF) is a critical component of rainfall redistribution, averagely contributing approximately four-fifths of gross rainfall into floor (Crockford & Richardson 2000;Levia & Frost 2006). Banana (Musa nana Lour.) plantations are widely cultivated in tropical and subtropical areas, where rainfall rates are high. The large broad-leaved canopy and abundant rainfall input dramatically modify the TF pattern under the banana canopy (Bassette & Bussière 2005). Considerable TF volume is generated at some dripping areas, six times greater than the incident rainfall rate under the banana canopy (Cattan et al. 2007;Yang et al. 2020). However, previous studies did not discuss TF variability in space and time scales in banana plantations, leaving some eco-hydrological problems unexplained, such as the reasons for heterogeneous soil water distribution and surface runoff rate (Klos et al. 2014;Zhu et al. 2021). Therefore, a thorough understanding of spatial-temporal variability of TF is essential in banana plantations (Yang et al. 2020).
The spatial heterogeneity of TF differs depending on the canopy structure (Staelens et al. 2006;Liu et al. 2019). For example, great TF volume appears in leaf dripping for the funnelling effect, and low TF volume appears for the shading effect (Germer et al. 2006). Some studies found that TF volume is related to the distance from tree trunks, e.g., greater TF either at crown edges (André et al. 2011) or near the trunk (Fan et al. 2015), but other studies found no such relation (Carlyle-Moses et al. 2004;Kowalska et al. 2016). Moreover, TF shows a consistent spatial pattern in high rainfall events and inconsistent spatial pattern in low rainfall events; this phenomenon can be attributed to the relatively large canopy interaction for small rainfall amounts and satisfied canopy water storage capacity for large rainfall amounts (Keim & Link 2018). Such spatial variability with increasing rainfall amount has been consistently documented in different forest systems, such as in shrub stands (Zhang et al. 2016), oak stands (Rodrigo & Ávila 2001), beech stand (Dezhban et al. 2019) and Douglas fir stands (Bouten et al. 1992). Nevertheless, these forest types still show different degrees of spatial variability of TF because of their different canopy structures and rainfall parameters. Therefore, further study of TF in various canopy types or typical climate regions is important to gain further insight into the spatial heterogeneity of TF (Levia et al. 2019).
In addition to spatial variability, the ranked plots of TF volume were used to assess the temporal variability of TF patterns over rainfall events (Keim et al. 2005;Liu et al. 2019). Generally, TF is characterised by a temporally variable pattern, considering that canopies are considerably diverse in terms of species and/or event scale. For instance, TF in a mature Japanese cypress (Chamaecyparis obtusa) stand was proved to have a slightly temporal stability because of canopy change (Sun et al. 2015). Differently, several studies reported that the TF pattern is characterised by considerable time stability related to stable canopy structure (Wullaert et al. 2009;Zheng et al. 2019). Sato et al. (2011) noted that a high temporal stability with the persistence of very wet positions in a eucalyptus plantation was related to the canopy persistently manifesting a funnel effect on TF. If spatially heterogeneous water inputs are temporally stable over a long time, the spatial distribution of soil water and solute deposition would be affected in the long term (Klos et al. 2014;Zheng et al. 2019). Therefore, a deep understanding of the temporal stability of TF patterns is crucial to estimate the ecological consequence of water input into floor.
Banana plantations are largely developed in the tropics and subtropics around the world, e.g., Central America, Asia and Africa. In China, banana plantations cover 331,900 ha, accounting for over 6% of the total banana plantation area of the world (CSY 2018). However, the abnormal ecological environment caused by large-scale banana cultivation induced Fusarium wilt and nematodes, resulting in yield loss (Almeida et al. 2018). The eco-hydrological problems (e.g., water and soil loss) of banana plantations have received minimal attention. In order to delve into the hydrological process of banana plantations, our team has carried out some substantial studies over the past few years. Fortunately, several insights have been gained, including the basic characteristics of TF volume and TF kinetic energy in banana plantations (Harris 1997;Zhang et al. 2021), runoff affected by rainfall redistribution (Cattan et al. 2009) and the efficiency of water use in banana crops (Yang et al. 2020). However, many important questions remain unanswered, such as the spatio-temporal variability of TF and raindrop size. Understanding the spatial and temporal patterns of TF is necessary to optimise the management of banana plantations in terms of soil water and nutrition availability. Moreover, TF drop, generated by leaf dripping points, was different from open rainfall. In detail, larger TF drops were generated by rainfall drops coalescing on leaf surfaces and appeared to be of variable size distribution resulting from rainfall and leaf shape (Brandt 1989;Nanko et al. 2006). Understanding the distribution of TF drop size would help in assessing soil splash erosion (Levia et al. 2019). Therefore, this study focused on the spatial heterogeneity, temporal stability and drop size of TF to integrate the hydrological process of banana plantations. The objectives of this study are to: (1) quantify the spatial heterogeneity of TF under banana canopy using variogram method; (2) identify the time stability of the TF patterns by several temporal levels; and (3) estimate TF drop size by using the filter paper technique.

Study site
The study site is located in Xishuangbanna Tropical Botanical Gardens (XTBG, 21.93°N, 101.27°E, 570-600 m a.s.l.), Yunnan Province, Southwest China. The local climate is dominated by a southwest monsoon carrying Indian Ocean moisture, and XTBG generates an apparent dry-wet season, including a rainy season from May to October and a dry season from November to April. According to the climate over the past 40 years, the mean annual air temperature and precipitation were 21.7°C and 1,480 mm, respectively; approximately 87% of the annual total precipitation occurs during the rainy season (May to October) and 13% during the dry season (November to April) (Liu et al. 2016).
The study was conducted in a 4.3-ha banana plantation comprising a banana monoculture planted at a density of 5,000/ha (tree spacing 2 Â 2 m). The stand characteristic was surveyed through 20 banana plants. In detail, the mean height of the mature banana plants was 295.1 cm, the mean basal area was 297.0 cm 2 , the mean diameter at breast height was 18.7 cm, and the mean leaf area index (LAI) was 3.7. Mature banana bunches were harvested between April and May at three growth stages: vegetative, flowering and bunch appearance. After harvest, the pseudostem was chopped down and the new banana seedlings that sprouted from bulbs grew rapidly. The banana plantation was properly irrigated to meet the demand for water during the dry season, and was supplied by rainfall during the rainy season. The banana plantation was managed through the frequent application of fertilisers (10-15 times a year) near the root of each plant, and regular weeding and pest control (pesticide spraying) were performed throughout the year.

Data collection
Rainfall and TF samples were collected from June to September 2019. Rainfall was collected using a tipping-bucket rain gauge (3554WD, Spectrum Technologies Inc., USA), which was installed in an open field adjacent to the banana plantation at a distance of about 300 m. Rainfall data were recorded using a data logger (Model 115, Spectrum Technologies Inc., USA). The resolution of the rain gauge was +0.2 mm, and the time interval of data recording was set to 10 min.
TF was collected using an artificial funnel-type collector, comprising a short-stemmed funnel (6.0 cm in diameter) and a polyethylene bottle (1 L in volume). Further, 32 collectors were installed under the banana plantation canopy, wherein each collector was located column-wise in the middle of two plants. The plot area was 8 m Â 16 m with eight rows (1-8) and four columns (A-D) at 2-m spacing. Moreover, three replications (open rainfall) were performed as control-experiment in the open field, 20 m from the banana field. TF volume (mm) was acquired from the water volume of the bottle divided by the horizontal cross-sectional area of the funnel.
The drop size distribution of TF at the dripping points was measured using filter paper staining (Brandt 1989). Dye composed of eosin and talc with a weight ratio of 1:10 was evenly coated on the filter paper of 20 cm diameter. The leaf shapes at the dripping points were classified as valley, overlap, tip, breakage and complex shapes (Zhang et al. 2021). The pits on the leaf margin are uniformly defined as 'valleys' (Va) in our study. The 'overlap' (Ov) of concave areas sometimes appears where raindrops are intercepted by the upper leaf, then accepted by the lower leaf and finally drained away. In fact, the Va and Ov shapes are both microcatchments on the leaves. Moreover, the leaf 'tip' (Ti) and the 'breakage' (Br) approximate the triangular geometrical morphology, in which the water is intercepted and drained away based on the slope. Finally, we classified leaves at canopy drip points into four shapes: Va, Ov, Ti and Br. Moreover, some leaf shapes such as Va-Ov, Ov-Br, Va-Br and even Va-Ov-Br were also observed, and we uniformly defined them as complex shapes (Cs). In this study, five leaves per shape and two replications per leaf were prepared. A device of artificial rainfall was used with steady spray intensity. Artificial rainfall was continuously applied. The filter paper was quickly moved to the ground point where the raindrop was falling; when caught in the raindrops, the filter paper was pulled out quickly to prevent overlap raindrop stains.
In addition, the leaf area index (LAI) of the banana canopy structure was measured using a LAI-2200 plant canopy analyser with 90°view caps (Li-Cor Inc., USA). LAI values were determined above every TF position once a month (four times in all).

Throughfall-related analysis
TF variability is dependent on rainfall event size (Gómez et al. 2002). To analyse TF variability systematically, we built a suitable TF classification based on certain parameters via Spearman rank correlation analysis and Pearson correlation analysis ( Table 1). The correlation coefficients related to the gross rainfall and the peak 30 min rainfall intensity (I 30 ) showed significantly higher values than other parameters. Thus, gross rainfall and I 30 were selected as the optimal abiotic parameters for TF classification. Thus, the recorded 24 rainfall events were classified by the rainfall amount (i.e., 0 to ,5, 5 to ,10, 10 to ,20, 20 to ,30, 30 to ,50 and !50 mm) and by I 30 (i.e., 0 to ,5, 5 to ,10, 10 to ,15, 15 to ,20 and !20 mm). The rainfall event classifications were further expressed as P1, P2, P3, P4 and P5 for the five gross rainfall classes and as I1, I2, I3, I4 and I5 for the five rainfall intensity classes. Descriptive statistics, including the accumulative TF and corresponding TF rate, coefficient of variation (CV, the standard deviation as a proportion of the mean), skewness and kurtosis, were calculated for all groups (Table 2).
Variogram is widely applied to quantify the spatial dependence among geographic elements (Isaaks & Scrivastava 1989). In this study, we calculated the variogram γ(h) to analyse the spatial autocorrelation of TF observations in the banana plantation.
The variogram γ(h) was computed as follows: where n(h) is the number of sample pairs, and the index i is the different spatial location of the samples. The values of samples z(x i ) and z(x iþh ) at two positions x i and x iþh are separated by the lag distance h. An appropriate semivariogram for the cumulative TF is computed by adjusting active lag distances until the minimum residual sum of squares (RSS) and the maximum coefficient of determination (r 2 ) appeared. The spatial autocorrelation is evaluated by the parameters of the model, nugget (C 0 ), sill (Cþ C 0 ), partial sill (C ). Relatively strong autocorrelation structure is defined by the partial sill/sill .75%; somewhat weak autocorrelation structure is defined by the partial sill/sill ,75%; the partial sill/sill close to 0% features a pure nugget structure and displays no spatial correlation (Keim et al. 2005). The fractal dimension, which is defined as the slope of γ(h), is used to ascertain the degree of spatial heterogeneity in different directions. The fractal dimension (D) was calculated as Note: Gr, gross rainfall (mm); I 10 -I 30 , peak 10, 20 and 30 min rainfall intensity (mm 10 min -1 , mm 20 min -1 and mm 30 min -1 ); Ave. I, mean rainfall amount per hour (mm h -1 ); ***p , 0.001, **p , 0.01, *p , 0.05. Note: P1-P5, five groups classified by gross rainfall amount; I1-I5, five groups classified by the peak 30 min rainfall intensity. follows: where m is the slope of linear regression. The values of D were computed in four directions: 0°(east-west), 45°(northeastsouthwest), 90°(south-north) and 135°(northwest-southeast).
To quantify the temporal persistence of the spatial distribution of TF, we normalised the TF volume by using the following equation from Keim et al. (2005): where T ij is the normalised TF at sampling point i for event j, T ij is the TF at sampling point i for event j and T j and SD j are the mean TF and standard deviation of the TF during that rainfall event, respectively. The normalised TF was ranked from minimum to maximum. This type of graph explained two types of temporal stability. First, the extreme persistence was defined as a large deviation of the normalised TF in the lower and upper quartiles of the arranged positions, where the sites are persistently very wet or dry at steeply sloping tails of the ranked line. Second, the general persistence was defined as a general deviation of the normalised TF in the interquartile range, where the sites are persistently wetter or drier than the mean TF (Keim et al. 2005).
For the drop size experiment, after allowing drying naturally, the filter paper was scanned with 1:1 into the computer and the stain diameter was calculated by software Image J. Previous studies have documented that a good correlation between stain diameter and raindrop diameter was found (Dou & Zhou 1982;Xu et al. 2004;Lu et al. 2015). Zhu et al. (2014) once studied the drop size distribution in this region and built an appropriate relation between stain diameter and raindrop diameter. Thus, the TF drop size diameter in this study was calculated using the following formula built by Zhu et al. (2014): where d is the actual drop size diameter (mm), and D is the diameter of stains on the filter paper (mm). The stain diameters for the drop size experiment were measured using Image-J software.

Statistical analysis and calculations
The normality of TF was tested by using the Kolmogorov-Smirnov test. A log-transformation or square root transformation was conducted for non-normally distributed data. The variograms of TF in different categories were calculated using GS þ 9.0 (Gamma Design Software). The differences in TF drop size among different leaf shapes were assessed using one-way analysis of variance (ANOVA), followed by a Tukey test for multiple comparisons between means. Spearman rank of correlation coefficients and Pearson's coefficients were used to test for linear correlations between the variables. All statistical procedures were conducted with α ¼ 0.05 threshold for significance, in the IBM SPSS statistics 22.0 software (IBM Inc.).

Rainfall and throughfall quantity
The total rainfall was 832.6 mm during the study period of 2019, which was lower than the annual mean value over the past decade (2009-2018) (Figure 1). The rainfall amount in the rainy season accounted for 77.4% of the total annual rainfall, similar to the annual record over the past decade. A total rainfall amount of 626.3 mm was collected for 24 rainfall events (0.9-69.7 mm) during the TF sampling period, with approximately 75.2% of the annual rainfall. The cumulative TF volume was 508.7 mm, on average 81.2% of the gross rainfall in the sampled banana plantation (Table 2). Moreover, descriptive statistics for each classified group showed that TF rate ranged within 77-86%, skewness À0.69-0.35 and kurtosis À1.18-0.55.

Spatial heterogeneity of throughfall
The spatial distribution of TF is heterogeneous for rainfall events (Figure 2). The number of wetter TF collectors beyond open rainfall increased with increasing rainfall volume and rainfall intensity. The distribution patterns of high-value areas and lowvalue areas were significantly different in rainfall groups, and there were more high-value points under heavy rainfall events (e.g., P4, P5, I4 and I5) than those under smaller rainfall events (e.g., P1, P2 and I1). For the spatial correlation, the best-fitting theoretical models for the TF semivariance via rainfall classification are shown in Figure 3. The fractural D values showed zonal anisotropy in different directions; therefore, the theoretical models were fitted. Some differences in the variogram results were found among different rainfall event classification (Table 3). In particular, in the gross rainfall classification, a spherical model fitted for two events (i.e., P1 and P2) showed a nearly stable sill in the range of 4-5 m (4.47 m and 4.34 m, respectively). Two events (i.e., P3 and P4) were fitted with a linear model, indicating an unclear stable sill. P5 was fitted with an exponential model for a nearly stable sill at the range of 3.3 m. By contrast, TF variance was fitted with higher proportions (.75%) in the P1, P2 and P5 groups compared with the P3 and P4 groups. In the rainfall intensity (I 30 ) classification, a spherical model and an exponential model best-fitted for two events (i.e., I1 and I5, respectively) showed a nearly stable sill in the range of 4-5 m. TF for I4 was fitted with an exponential model for weak correlation. However, the other events (i.e., I2 and I3) were fitted with a linear model, suggesting unclear stable sills and low proportions (both 0%). The CVs of TF are scattered with gross rainfall and I 30 in Figure 4. TF in the banana plantation showed spatial variability mostly with CVs . 20%, especially at some points .50%. Notably, the CVs of TF varied with increasing gross rainfall and I 30 , but extremely weak nonlinear relationships were fitted between CVs and gross rainfall and I 30 . Thus, their relationships were fitted by partitioning two different zones (small and heavy rainfall events). The CVs significantly decreased by exponential function for both zones (R 2 . 0.5, p , 0.001) (Figure 4(a)). In the transitional zone (green pane filled with dots), decreasing CVs for small rainfall events (,15 mm) increased, and then high CVs decreased again with increasing gross rainfall for events .30 mm. In addition, the CVs significantly decreased with I 30 by exponential function for both zones (R 2 . 0.6, p , 0.001) (Figure 4(b)). Meanwhile, the transitional zone ranged from 10 mm 30 min À1 to 15 mm 30 min À1 . In addition, in different classifications for each group, the CVs ranged from 16% to 35% (Table 2). Meanwhile, the CVs initially increased and then decreased from P1 (and I1) group to P5 (and I5) group, with peak values appearing for the moderate rainfall events (i.e., P3 and I3 groups). The spatial distribution of TF showed a higher variability in the P3 group, with a CV of 34.5%, than those in other groups. Similar results of TF variability were also found in I 30 classifications for each group.  The temporal stability of spatial TF patterns in the banana plantation was studied at three temporal levels based on Spearman rank correlation. At the first temporal level, relationships among the cumulative TF volumes across different rainfall classes P1-P5 were determined (Table 4). The significant correlation (p , 0.05) among the cumulative TF volumes per rainfall class was not fully determined and was found only in P2, P4 and P5. At the second temporal level, relationships among the monthly cumulative TF volumes across four months (June-September) were determined. The monthly cumulative TF volumes showed no significant or strong correlation, except the correlation between August and September (r ¼ 0.37, p , 0.05). At the third temporal level, the correlations with each other of the TF volumes for all rainfall events were calculated, and a significant (p , 0.05, n ¼ 32) positive Spearman correlation coefficient was found for 10.9% of the pairs of events (numbers of pairs of 24 rainfall events ¼ 276) (Table 4). Theoretical model, nugget, sill, R 2 , RSS and ratio between partial sill and sill parameters to percentage can be seen. Note: P1-P5, five groups classified by gross rainfall amount; I1-I5, five groups classified by the peak 30 min rainfall intensity. In addition, the temporal stability plots for the mean normalised TF in each collector position and for different rows are shown in Figure 5. For total rainfall events, the mean normalised TF was significantly different than zero for 34.4% of the samples (t-test; α ¼ 0.05). No TF collectors were very dry, and a tail indicated persistence of individually wetter TF collectors, showing moderate general persistence in the stand. For each row, a moderate general persistence was also found, with no rows that were extremely wet or dry. Moreover, each plot of four months was ranked in ascending order (Figure 6), and results showed that TF was characterised by great general persistence, particularly the persistence of extremely dry collectors (e.g., 3C, 6C, 7D and 7D corresponding to June, July, August and September, respectively) and few very wet collectors (e.g., 1B, 2B, 7A and 4A corresponding to June, July, August and September, respectively).

Drop size distribution of throughfall
The drop size of TF was in the range of 3-10 mm and was divided into different proportions in each leaf shape (Figure 7). For the Va shape, the number of TF drops in the range of 3-4 and 9-10 mm showed larger proportions of 33.3% and 35.0% of the number of total raindrops, respectively, than other drop sizes. The TF drop size for the Ov shape was mainly in the range of 8-10 mm, a proportion of 47.0%. A large proportion of drop size ranging from 7 to 10 mm was found for the Br shape, which accounted for 93.7% of the total drops. The TF drop size for the Ti shape was mainly in the range of 8-10 mm with a proportion of 80.0%. As for the Cs shape, a TF drop size of 8-9 mm showed a larger proportion of 46.1%, than other drop  sizes. In addition, a significant difference (p , 0.05) in TF drop size was found among the different leaf shapes. On average, the drop size of TF followed a decreasing order of Br . Ti . Cs . Ov . Va.

Comparison of spatial heterogeneity in throughfall
The spatial dependence of TF in a certain range reflected the spatial heterogeneity of TF because of the interaction of the surrounding TF plots (Gómez et al. 2002). In this study, the TF was spatially correlated up to a range of approximately 3-5 m and presented a strong autocorrelation structure. This suggested that a sampling distance could be considered .5 m in the banana plantation. Gómez et al. (2002) found a nearly stable sill at an approximated 1-3 m distance in isolated olive trees. Staelens et al. (2006)  trees, suggesting that larger sampling distance should be conducted in the oak trees. These differences in the spatial heterogeneity of TF are expected, and one reason is that the differences in canopy species, canopy structure (e.g., tree height, leaf area and LAI) and meteorological phenomena alter the spatial patterns of TF (Keim & Link 2018). The banana tree with about 3 m height is generally lower than most forest trees, and the banana leaf area (about 1.5 m 2 per leaf) is significantly larger than that of the common vegetation leaf, all of which cause the distinct different distribution pattern of TF. Another reason is that the choice of sampling schemes, such as area extent, sampling size and layout, has a substantial influence on the variogram results (Voss et al. 2016). Besides, groups P4, I2 and I3 had extremely small sill proportions (0.0%), displaying no spatial correlation. Similar to our finding, the spatial correlation of TF was absent in a maritime pine stand (Loustau et al. 1992) and in Mediterranean dry forest (Bellot & Escarre 1998). The variograms in the I2 and I3 groups did not reach a stable sill at a range possibly because of the high minimum intermediate distance between the TF collectors, resulting in a lack of information on the spatial correlation at a small distance (Fathizadeh et al. 2014). The spatial variability of TF was verified to significantly decline with increasing rainfall amount in previous researches (Whelan & Anderson 1996;Fan et al. 2015). By contrast, our study found weak correlations between CVs and rainfall amount and I 30 for the 24 monitored rainfall events ( Figure 4). Fortunately, we found significant correlations when rainfall events were divided into two intervals, relatively low and high rainfall events, respectively. Results showed that two distinctly exponential trends were observed with increasing rainfall amount and intensity. The two forces from rainfall and canopy intensively counteract each other and induced such results. In particular, for small rainfall events, canopy interactions are relatively large, resulting in a high degree of spatial TF variability initially (Lin et al. 1997) and a strong correlation between CVs and rainfall amount. With the increase of rainfall amount, the splash intensity of dripping water on the leaves increases, which induces canopy shaking. The shaking canopy as well as the converging and shading effects of the leaves interfere with the falling pattern of raindrops. Thus, for the moderate rainfall events, the CVs were unsteady likely due to a transitional interaction between rainfall and canopy factors. For the high rainfall events (.30 mm), the CVs again declined with increasing rainfall amount. Moreover, the CVs in the P4, P5 and I5 groups were smaller than those in the other groups (Table 2), also indicating that a relatively low degree of TF spatial variability would be generated in high rainfall events. Similarly, previous studies also demonstrated that the spatial variability of TF was restricted by canopy complexity and rainfall depth (Lloyd & Marques 1988;Holwerda et al. 2006;Kowalska et al. 2016). Carlyle-Moses et al. (2004) noted that CV value decreased with increasing rainfall volume and afterwards remained quasi-constant after a rainfall threshold was reached. Their study suggested that the spatial distribution of TF volume was also determined by the distribution of canopy drip and shelter point locations. In banana plantations, unexceptionally, the spatial variability of TF can be attributed to these documented factors, rainfall and canopy structure. In summary, the increasing rainfall volume induced the decrease in TF spatial variability for the 24 monitored rainfall events. Considering the extremely long, wide banana leaves, the spatial distribution of TF was likely dependent on the banana canopy structure, particularly the dramatic variability in the TF volume due to the banana canopy funnelling and shading effects.

Temporal stability of throughfall patterns
In this study, the significant correlation of TF volume was not fully caught at the temporal levels based on Spearman rank correlation, indicating that TF pattern was inconsistent across rainfall events in the banana plantation. At the experiment scene, we noticed that the dry TF collectors were mostly adjacent to the shade of banana leaves, a few of which never received any raindrops. The wet collectors were obviously related to the convergence effect of large banana leaves. In general, the shading and funnelling areas for the canopy were distributed at random. Thus, the case of persistently dry and/or wet TF in the same collectors rarely appeared across rainfall events. Moreover, the time stability of spatial TF pattern was lower (10.9%) at the level of individual rain events compared with the results (14.2%-28.8%) in individual oak trees (Quercus brantii var. Persica) reported by Fathizadeh et al. (2014). This difference suggests that the canopy factor control on TF variability may be more complex than the rainfall parameter control. In addition, the proportion of sampling points (approximately 34%) deviating consistently in two directions from the mean normalised TF in banana plantation was similar to the 31-46% detected in two coniferous forests and a deciduous forest (Keim et al. 2005) but higher than the 1.4%-5.2% found in Persian oak trees (Quercus brantii var. Persica) (Fathizadeh et al. 2014). Thus, our result illustrated that TF volumes at sampling points were not entirely randomly distributed over time, and the TF volume obviously showed a high temporal variability. Similarly, Germer et al. (2006) also reported a high temporal variability in an open tropical rain forest. However, there is no consistence among previous studies; for instance, Keim et al. (2005) found a low temporal variability for coniferous and deciduous stands. As described above, a possible explanation for the low temporal stability (namely, high temporal variability) of TF spatial pattern in our experiment should be attributed to frequent changes in canopy properties (Loescher et al. 2002).
In addition, the correlation coefficients among TF, GR and I 30 for each TF plot were calculated to identify the effect of rainfall events on TF from 32 different collectors. Results showed that the quantity of collectors with significant correlation (p , 0.05) was obviously less than the collectors with weak and even no correlation (Supplementary Material, Figure S1). As previously stated, the differences among TF collectors in terms of the effect of rainfall characteristics are potentially a result of biotic factors, among which, canopy structure is a major contributing factor to rainfall redistribution (Zimmermann et al. 2007;André et al. 2011). LAI has been proved to exert an important effect on the TF variation of other forest systems (Song et al. 2018). In this study, LAI values under the banana canopy were measured to verify LAI's effect on TF variability. Unfortunately, significant correlation between TF volume and LAI was not found. The raindrop funnelling and shading areas for the canopy were distributed at random around the banana plantation, thus, not affected by the high or low LAI. In fact, the funnelling and shading areas of banana canopy are promoted by the leaf size, whose area is often up to 1.5 m 2 , much greater than that of common tree leaves. Thus, the effect of the banana canopy on rainfall redistribution was different from that of other forest canopies (Staelens et al. 2006). Overall, the spatial-temporal variability of TF in the banana plantation showed peculiarity resulting from the large broad-leaved canopy.

Difference in throughfall drop size
In addition to promoting the pronounced spatial-temporal variability of TF, canopy effect on TF was also reflected in various components of TF drop size. In general, large TF drops are generated by rainfall drops coalescing on the leaf surfaces (Nanko et al. 2006). In the banana planation, the TF drop size distribution of 3-10 mm showed a great range of canopy drip. However, previous studies reported that the TF drop size distribution of canopy drip is mostly ,7 mm (Hall 2003;Levia et al. 2019), which is significantly lower than the peak of banana canopy drip. Drop size of tall trees was found to be lower than that of short shrubs in Amazon forest systems, indicating that tree height was an essential factor affecting leaf drop size (Rebelo & Williamson 1996). Therefore, the result, relatively large drop size of banana leaf, can be attributed to relatively low height compared to tall trees. In addition, the large drop size (7-10 mm) of the banana canopy drip can be ascribed to the fact that the extremely large leaf size facilitated the coalescence of raindrops on the leaf surface. Such case is no exception, and some studies have indicated that the broad-leaved tree (or matte leaves) indeed could generate larger TF drop size than the coniferous tree (or coated leaves) (Nanko et al. 2006(Nanko et al. , 2013. Significant differences in TF drop size distribution were observed among different leaf shapes. This result indicated that the ability of the raindrops to coalesce in different leaf shapes might be a result of such significant differences. Accordingly, specific canopy with different leaf shapes could be considered as a key biotic factor affecting the TF volume and drop size distribution of certain species.

Implications
The spatially heterogeneous TF input produces wetter and drier soil areas within a short time (Zheng et al. 2019). Soil subsurface flow or rapid recharge to groundwater can be affected by soil surface water distribution (Weiler & Naef 2003). In the sampled banana plantation, the highly spatial variability of TF was detected. As a result, heterogeneous TF input is likely to produce extremely wet and extremely dry soil areas. Thus, whether or not the high spatial variability of TF contributes to some differences in soil water dynamics in the banana plantation is worthy of further study in future.
Heavy rainfall can provide abundant water for soil moisture and also leach canopy elements to soil solute concentrations. However, the negative effect is that heavy rainfall may generate serious water and soil loss via surface runoff. In our study site, soil erosion frequently occurs during the rainy season, which covers six months each year (Figure 8). The banana canopy converges raindrops and TF shows large temporal and spatial variability, which affects the soil splash erosion distribution. Some studies have reported that the TF splash erosion potential in agroforestry systems is significantly lower than that in monocultures because multiple-layered canopies can prevent the raindrop size and terminal velocity of intercropping plants Liu et al. 2018). In practice, the agroforestry ecosystem of multiple canopies can be considered instead of the monoculture banana plantation.

CONCLUSION
This study confirmed the spatial heterogeneity and temporal stability of TF for 24 rainfall events and the TF drop size in a banana plantation. The TF volume showed strong spatial correlation up to a range of approximately 3-5 m. The spatial pattern of TF maintained high variability, and the spatial variability twice exponentially declined with increasing rainfall amount in two separate rainfall intervals, respectively. This unusual trend can be attributed to the significant canopy effect against rainfall effect in different ranges of rainfall event size. Moreover, the observed spatial pattern of TF was temporally low and stable during the four months of observation, which was controlled by an unstable banana canopy structure. In addition, the TF drop size from the leaf dripping points varied highly in 3-10 mm and had significant differences (p , 0.05) among five types of leaf shapes. Thus, large broad-leaved canopy of banana can dramatically modify the spatial-temporal variability of TF.