Stability of Soil Moisture Patterns Retrieved at Different Temporal Resolutions in a Tropical Watershed

Above and underground hydrological processes depend on soil moisture (SM) variability, driven by different environmental factors that seldom are well-monitored, leading to a misunderstanding of soil water temporal patterns. This study investigated the stability of the SM temporal dynamics to different monitoring temporal resolutions around the border between two soil types in a tropical watershed. Four locations were instrumented in a small-scale watershed (5.84 km) within the tropical coast of Northeast Brazil, encompassing different soil types (Espodossolo Humilúvico or Carbic Podzol, and Argissolo Vermelho-Amarelo or Haplic Acrisol), land covers (Atlantic Forest, bush vegetation, and grassland) and topographies (flat and moderate slope). The SM was monitored at a temporal resolution of one hour along the 2013-2014 hydrological year and then resampled a resolutions of 6 h, 12 h, 1 day, 2 days, 4 days, 7 days, and 15 days. Descriptive statistics, temporal variability, time-stability ranking, and hierarchical clustering revealed uneven associations among SM time components. The results show that the time-invariant component ruled SM temporal variability over the time-varying parcel, either at high or low temporal resolutions. Time-steps longer than 2 days affected the mean statistical metrics of the SM time-variant parcel. Additionally, SM at downstream and upstream sites behaved differently, suggesting that the temporal mean was regulated by steady soil properties (slope, restrictive layer, and soil texture), whereas their temporal anomalies were driven by climate (rainfall) and hydrogeological (groundwater level) factors. Therefore, it is concluded that around the border between tropical soil types, the distinct behaviour of time-variant and time-invariant components of SM time series reflects different combinations of their soil properties.


INTRODUCTION
Soil moisture (SM) is a crucial variable affecting biological, hydrological, and biogeochemical processes (Liu et al., 2010;Coppola et al., 2011;Rossato et al., 2013). However, the heterogeneous aspects of environmental factors convey an inherent high variability in time and space to SM, ranging from minutes to years and from centimetres to several kilometres . Some of the main factors driving SM variability are soil texture, vegetation, topography, and meteorological conditions (Mittelbach and Seneviratne, 2012;Liu et al., 2015), whose interactions trigger complex effects on SM time series (Zhu et al., 2014), which calls for a comprehensive analysis of its temporal stability and dynamics (Zhou et al., 2015). On this topic, an increasing number of studies have recently addressed new approaches to test SM spatiotemporal variability (Robinson et al., 2008). For example, Vereecken et al. (2014) have reviewed methods and techniques used to characterise SM at the field scale, including recent measurement advances such as remote sensing. Other studies have applied some of these techniques to analyse SM in small-, medium-, and large-scale watersheds under a wide variety of hydrological and climatic conditions (Vachaud et al., 1985;Qiu et al., 2001;Brocca et al., 2010;Souza et al., 2011;Molina et al., 2014;Zhu et al., 2014;Liu et al., 2015;Sun et al., 2015;Huang et al., 2016). Such analyses encompass techniques that hold copious applications that allow for revisiting the standard soil sampling schemes (Souza et al., 2011) and resetting the default monitoring time-step (Molina et al., 2014).
The SM stability concept, from a temporal perspective, was first introduced by Vachaud et al. (1985) to address time-invariant point-scale characteristics. The authors found a considerable time-stability in some locations, representing mean and extreme behaviours of SM throughout the year. Mittelbach and Seneviratne (2012) have decomposed a long-term SM time series between the temporal mean and temporal anomaly parcels to weigh the extent of that concept. Their results highlight the applicability of the invariant component in comparison to small contributions of the dynamic component to the overall variability. Conversely, other studies have found considerable implications of temporal anomalies on climate modelling (Viola et al., 2008), sub-seasonal forecasting (Weisheimer et al., 2011), land-atmosphere interactions (Hang et al., 2016), and satellite-derived product validations (Rötzer et al., 2015). The reliability of the stability concept is also intimately affected by the time-step of SM monitoring, because as the probe coverage footprint decreases, any monitoring disturbance tends to alter its mean integrated estimate. Molina et al. (2014) have claimed that a 10 × 10-m mesh grid (nine probes) would allow for estimating the mean time-stable SM in small homogeneous parcels. The authors reported no loss of information upon the daily SM variability by decreasing the 20-min temporal resolution to eight hours and one day during wetting-up and drying-down periods, respectively. In this context, temporal SM anomalies are of great interest in spatial variability analysis because they reduce the impacts of static soil properties on time series, although a considerable fraction of the spatial variability of these anomalies may be time-invariant (Mittelbach and Seneviratne, 2012;Brocca et al., 2014).
Despite the considerable number of SM spatiotemporal studies, some authors encourage further investigations over their time-variant and -invariant components worldwide due to the incipient knowledge on some ungauged regions (Mittelbach and Seneviratne, 2012). Brocca et al. (2014) have studied six different SM monitoring networks in distinct geographical regions and found that four featured standard deviation values of temporal anomalies lower than the standard deviation of raw SM data. Moreover, half of the addressed networks showed minimal spatial variability in moderate SM conditions when considering temporal anomalies. Souza et al. (2011) have applied the SM-relative differences ranking to identify sites in the Brazilian semi-arid that could reproduce the mean SM spatial pattern for agricultural water management purposes, while Silva Junior et al. (2016), Melo and Montenegro (2015), and Silva et al. (2015) have investigated the SM temporal dynamics and verified the importance of SM temporal stability studies for reducing hydrological monitoring costs under different land cover and soil types. Concerning the impact of SM temporal dynamics, Costa et al. (2016) have investigated the permanence of soil water effectiveness under preserved-vegetation constraints in three semi-arid soil types. In a tropical volcanic archipelago, Montenegro et al. (2009Montenegro et al. ( , 2018 have investigated the water potential related to SM dynamics in three soil types to support local soil water management and conservation practices. Other relevant studies have also recently been carried out to assess the influence of SM dynamics on sedimentological processes (e.g. Montenegro and Ragab, 2010;Montenegro et al., 2013;Figueiredo et al., 2016). Table 1 provides a tabular chronological literature review of relevant studies on SM dynamics, also describing the advancements on the SM time-stability method.
All studies listed in table 1 have been carried out using the SM time-stability method applied in distinct areas throughout the world, at different spatiotemporal resolutions and with various methods of analysis. Nevertheless, Brocca et al. (2017) point out the scarcity of SM data available in the International Soil Moisture Network: relevant data are mostly from the USA and Europe. In Brazil, studies about spatiotemporal stability methods still prompt an insufficient understanding of regional SM heterogeneities in this continental-scale country, not allowing for extending conclusions of previous investigations to other particular soil types. Furthermore, the studies listed in table 1 provide insights into the effects of different temporal resolutions of monitoring over time-variant and -invariant components by neglecting the influences of SM peaks and recession mischaracterisation on the statistical metrics that summarise its temporal variability/stability. It is hypothesised that comparing the mean statistical metrics of the SM time-variant and time-invariant components at different temporal resolutions may allow for pointing out the temporal stability of soil water patterns.
Against this background, this study aimed to investigate the stability of the SM temporal dynamics (Mittelbach and Seneviratne, 2012) to different monitoring temporal resolutions (such as Molina et al., 2014) around the border between two soil types in a tropical watershed. For this purpose, the SM time series of four monitoring locations were assessed at eight temporal resolutions to identify how their time-variant and time-invariant parcels prompt SM temporal variability among sites with different soil types (Espodossolo Humilúvico or Carbic Podzol, and Argissolo Vermelho-Amarelo or Haplic Acrisol), land covers (Atlantic Forest, bush vegetation, and grassland) and topographic (flat and moderately sloped) features. The present study revisits and employs a framework of statistical metrics recently updated (Mittelbach and Seneviratne, 2012) for analysing the SM temporal dynamics on a point-scale basis, around a low-density and poorly-monitored tropical area with well-distinguished rainy and dry seasons.

Study area
The study was carried out at the Guaraíra Experimental Watershed (GEW) on the tropical coast of Northeast Brazil (Figure 1). The watershed covers 5.84 km 2 between the coordinates 7° 32'-7° 27' S and 34° 04'-35° 02' W. Since 2003, it has been operated and monitored by the Hydrology Network of Semi-Arid (REHISA), which is currently composed of eight federal universities in Brazil (Montenegro and Ragab, 2010;Coelho et al., 2017). The monitoring program set to this experimental watershed makes an important experimental contribution to Brazilian hydrology, since the GEW is strategically located near the transition zone between the Atlantic Forest (tropical climate) and Caatinga (semi-arid climate) biomes, in a coastal sedimentary area, where the largest cities in tropical Northeast Brazil are located. Its natural vegetation encompasses a mix of Capoeira (secondary-growth bush vegetation) and Atlantic Forest remnants; part of this native vegetation was also cleared for cultivating sugarcane and pineapple crops (Sales et al., 2014).  Vachaud et al. (1985) (i) Grenoble, France; (ii) Seville, Spain and (iii) Mornag, Tunisia/(i) n.r.; (ii) n.r.; (iii) n.r. / (i) 2 km 2 ; (ii) n.r.; (iii) 0.01 km 2 (i) 14 days and bimonthly; (ii) n.r.; and (iii) n.r./(i) 2.5 years; (ii) n.r.; and (iii) n.r./(i) neutron tubes; (ii) neutron tubes and tensiometers; and (iii) neutron tubes and gravimetric method Introduced the concept of soil moisture time-stability and variability in order to reduce the number of soil sample observations Qiu et al. (2001) Losses Plateau, China/Continental semi-arid/3.5 km 2 Biweekly/9 months/TDR Used the metrics proposed by Vachaud et al. (1985), associated with the local features (climate, soil type, and elevation), in order to reduce the number of local soil moisture measurements Brocca et al. (2010) Vallaccia catchment, central Italy/Mediterranean semihumid/56 km 2 35 sampling days with almost weekly frequency, except for the summer period/1 year/TDR Applied the methodology proposed by Vachaud et al. (1985) in distinct sites inside the catchment to define the minimal number of measurement points to validate soil moisture data provided by remote sensing products Souza et al. (2011) Mimoso catchment, North-Eastern Brazil/Semi-Arid very hot/1,800 m 2 Twice per week/96 days/ neutron tubes Evaluated the soil moisture time-stability concept proposed by Vachaud et al. (1985) and its dependence in an irrigated carrot plot Brocca et al. (2012) Genna and Caina catchments, central Italy/Mediterranean semi-humid/242 and 178 km 2 , respectively 35 and 34 sampling days with almost weekly frequency, respectively, except for the summer period/11 months/TDR Applied the methodology proposed by Vachaud et al. (1985) at small and medium spatial scales in order to reduce the number of measuring points across the region Mittelbach and Seneviratne (2012) 14 grassland sites in Switzerland/Six distinct climates/~31,500 km 2 Daily/14 months/TDR and capacitance Updated the method proposed by Vachaud et al. (1985), using a framework to distinguish the timevariant and -invariant contributors to spatial variability at country scale Brocca et al. (2014) Italy, Spain, France, Switzerland, Australia, and United States/Distinct climates/From 250 to 150,000 km 2 From daily to 15 days/ From 10 months to less than 3.5 years/n.r.
Compared the framework carried out by Mittelbach and Seneviratne (2012)  According to Köppen's climate classification system (Köppen, 1936), analysed in detail across Brazil by Alvares et al. (2013), the region holds a tropical climate with a dry summer (As), whose mean temperature is around 26 °C. Mean annual rainfall is ~1,700 mm, most of which (~70 %) occurs from March to July, i.e. in autumn-winter seasons . Accordingly, mean annual class-A pan evaporation reaches ~1,300 mm in the GEW (Silva et al., 2000). Its elevation ranges from 37 to 135 m above mean sea level, and its mean slope rate is 1.8 %, being more sloped at downstream and flat at upstream. Quaternary-age sediments, originating from regional basement erosion, cover the study area. Two soil types are present in the GEW: Carbic Podzol and Haplic Acrisol, according to the World Reference Base (WBR) for soil resources (IUSS Working Group WRB, 2015) ( Figure 1b). These soils are equivalents to Espodossolo Humilúvico Hidromórfico dúrico-arênico and Argissolo Vermelho-Amarelo Distrófico típico, according to the Brazilian Soil Classification System (SiBCS), respectively (Santos et al., 2018). Argissolo (Acrisol) holds a thick layer with deep well-drained soil, while Espodossolo (Podzol) presents an illuvial silica-cemented restrictive layer at a depth of ~2 m (MA and SUDENE, 1972). These features confer to the study area a complex underground system characterized by thickness alternations of the vadose zone, varying from thin to thick.
Near-surface SM measurements were registered at four fixed sites located within the GEW, monitoring rainfall (sites 1, 2, and 3) and meteorological data (site 4). The meteorological data available at site 4 were as follows: relative air humidity, soil heat flux, soil and air temperature, net radiation, and wind speed. The physical properties of each site are shown in table 2.

Soil moisture monitoring and temporal resolutions
Time-Domain Reflectometry (TDR) probes (CS616, Campbell Scientific) were installed, calibrated, and automatically monitored the SM of each site at one-hour temporal resolution. The TDR probes were vertically inserted to indirectly represent the integrated mean of SM along the first 0.30 m below the surface. The SM data were assimilated from April 2013 to April 2014, throughout the hydrological year. The original time-step of monitoring was then used as the reference scenario for this study.
Seven SM scenarios were investigated with increasingly coarser temporal resolutions of six h, 12 h, 1 day, 2 days, 4 days, 7 days and 15 days. Coarser temporal resolutions were obtained from the one-hour reference scenario by excluding the intermediary registers, e.g. the six-hour temporal resolution was obtained by ignoring the five previous hourly values. This procedure was the same as that used by Molina et al. (2014), hypothesising that the intermediate records are not available for the excluded time scales. The new temporal series were analysed and compared independently for each analysis described in the following sections.

Soil moisture and its time-variant and time-invariant components
In this study, S represents the SM, while time and space are termed t and n, respectively. Therefore, S tn indicates the SM at t and n. The variables T and N denote the total number of time registers and sites (four sites), respectively.
Similar to Mittelbach and Seneviratne (2012), the SM temporal mean is termed as M tn , the time-invariant component, whereas the SM temporal anomaly as A tn , the time-variant component. Thus, the relation between them is defined by equation 1: The SM variability and stability in time at any site were studied in the GEW through the combination of classical approaches (Vachaud et al., 1985) and improvements Mittelbach and Seneviratne, 2012;Brocca et al., 2014). Accordingly, the same notations used by Mittelbach and Seneviratne (2012) and Brocca et al. (2014) for the mean (μ) and standard deviation (σ) were replicated in this study. Then, equations 2 and 3 show the temporal mean μ t̂( S tn ) and the temporal standard deviation σ t̂( S tn ) at any site in the GEW, where the subscript t̂ was used for indicating the temporal statistic meaning, as follows: The absolute values of A tn , |A tn | were used to circumvent compensation between positive and negative values in the statistical metrics. Subsequently, equation 4 shows their temporal mean μ t̂( |A tn |), as follows: Furthermore, equation 5 shows the SM spatial mean μ n (S tn ) at any time, calculated as the arithmetic mean of S tn among the four sites. Similarly, equation 6 shows the |A tn | spatial mean μ n (|A tn |) at any time, calculated as the arithmetic mean of |A tn | among the four sites. The subscript n (without ^) was used here to indicate that this spatial mean does not represent the mean value within an area, but a mean value of spatially different sites:

Time-stability ranking upon the absolute and relative differences of time-variant and time-invariant components
The time-stability concept proposed by Vachaud et al. (1985) was adopted to identify the more representative sites to the SM spatial mean within a network. The notations Δ and δ, adopted by Vachaud et al. (1985) and Mittelbach and Seneviratne (2012), were used to represent the absolute and relative differences, respectively. Accordingly, the symbol ΔS tn is defined as the absolute difference between S tn and the SM spatial mean μ n (S tn ), as shown in equation 7: ΔS tn = S tn -μ n (S tn ) Eq. 7 Equations 8 and 9 show how the ΔS tn temporal mean μ t̂( ΔS tn ) and the standard deviation σ t̂( ΔS tn ) are calculated: Equations 10, 11, and 12 show how the relative difference δS tn and its temporal mean μ t̂( δS tn ) and standard deviation σ t̂( δS tn ) are calculated: Time-stability ranking was performed over the metrics obtained at different temporal resolutions for the four monitored sites, ordering their respective deviation from the overall mean. As such, the μ t̂( ΔS tn ) and μ t̂( δS tn ) from each monitored site in the GEW were arranged from the smallest to the largest. The closest points to μ n (S tn ), i.e. sites with Rev Bras Cienc Solo 2019;43:e0180236 μ t̂( δS tn ) and μ t̂( ΔS tn ) near zero, were then considered the most stable. Additionally, the structure proposed by Vachaud et al. (1985) and adapted by Mittelbach and Seneviratne (2012), which considers the absolute terms |ΔS tn | and |δS tn |, was also adopted. Their means and standard deviations were then similarly obtained and arranged in ascending order by their respective means, i.e. |μ t̂( ΔS tn )| and |μ t̂( δS tn )|.
The time-stability analyses carried out according to equation 7 to equation 12 can be extended through equation 1 to include the different components of absolute SM, as previously applied by Mittelbach and Seneviratne (2012). The absolute values of these components are used because A tn may have negative values. Consequently, the following equations are used: |ΔA tn | = |A tn -μ n (A tn )| Eq. 13 |ΔM tn | = |M tn -μ n (M tn )| Eq. 14 The temporal means and standard deviations of the absolute anomalies, i.e. μ t̂( |ΔA tn |) and σ t̂( |ΔA tn |), were ranked according to their respective deviation from the overall mean. Similarly, the absolute variation of the temporal mean ΔM tn results, i.e. μ t̂( |ΔM tn |), were also ranked. Lastly, comparison between the ranking of the temporal mean of absolute SM differences (|μ t̂( ΔS tn )|) and its decomposed parts (|μ t̂( ΔM tn )| and |μ t̂( ΔA tn )|), obtained at different temporal resolutions, was performed, which allow for inferring on how the time-variant and -invariant components (Vachaud et al., 1985) incorporate the SM temporal dynamics at different sites.

Descriptive statistics and hierarchical clustering
The descriptive statistics of the SM datasets at each site n and time t (S tn ), namely, the mean, standard deviation, coefficient of variation, minimum and maximum values, were calculated for every preselected temporal resolution: the reference scenario and the others seven scenarios (see section 2.2). Box plots were also drawn representing mean, median, quartiles (25 and 75 %), quantiles (10 and 90 %), and outliers (less than 10 %-and higher than 90 %-quantiles). Based on this, it was possible to identify how data losses may affect the temporal behaviour of the time series as well as which sites are more sensitive to a decrease in temporal resolution.
Hierarchical cluster analysis was also applied to the main time-variant and time-invariant components of the SM, i.e. S tn , ΔS tn , δS tn , A tn , ΔA tn , M tn , and ΔM tn , to assess the association between these variables. This analysis was used to compare how the dendrograms changed from site to site and within each site, considering different temporal resolutions. The Euclidean distance measure (Deza and Deza, 2009) and Ward's method (Ward Jr, 1963) were used to link the dendrograms. As different variables with unequal ranges were used in this study, the first step of the clustering comprised the procedure of standardisation, so that all of them ranged from -1 to 1. This standardisation was carried out through the mean and standard deviation of each variable dataset. Subsequently, pairwise distances were calculated using the Euclidean distance, termed as D link . The maximum distance was estimated between variables, termed as D max . This study used the relative Euclidian distance D link /D max (in percentage) to link the SM components and to enable comparison between different sites and different temporal resolutions at the same site.

Temporal analysis of absolute soil moisture
Approximately 30,000 SM measurements at one-hour temporal resolution were processed at each analysed site in the GEW throughout the study year 2013/2014. Some sites exhibited a lack of data throughout this period, which represents ~12 % of the total dataset. The accumulated rainfall over the area was 1,942 mm in this hydrological year, ~12 % higher than the mean long-term rainfall. Due to these forcing events, the SM time series displayed successive peaks and recessions throughout the hydrological year, in which the sites 2 and 3 featured greater and smaller amplitudes, respectively ( Figure 2). Rainfall events registered from June to September 2013 led to noticeable SM increases, whose values at all sites ranged from 8.1 (site 3) to 114.6 mm (site 2) for these months. The minimum SM values for Espodossolo Humilúvico (Carbic Podzol) were 5.6 (site 2) and 6.5 mm (site 3), whereas at Argissolo Vermelho-Amarelo (Haplic Acrisol), the values were 8.4 (site 1) and 12.2 mm (site 4). The SM temporal pattern during the summer dry season (from October 2013 to January 2014) presented maximum values up to 54.6 mm when all sites were considered.
The descriptive statistics of SM, used to analyse time periods according to the temporal resolutions in this study, is shown in table 3. Noticeably, since the temporal resolution is longer, the mean, standard deviation and coefficient of variation of SM values slightly increase. Regarding the SM maximum values, changes are more perceptible, decreasing from 33 to 42 % between temporal resolutions of one hour and 15 days at sites 1 and 4, respectively. Relevant changes in the maximum values of these two specific cases were observed between temporal resolutions of 12 hours and 1 day. Although more noticeable at sites 1 and 4, such decreasing behaviour implies some substantial loss of information also at site 2 with coarsening of the temporal resolution. However, although site 2 presented a higher peak oscillation (Figure 2), data losses were irrelevant for the temporal mean. On the other hand, non-extensive data losses were registered at site 3 due to its slow SM recession. Figure 3 presents the box plots of SM for every monitored site and all temporal resolutions analysed in this study. Visually, at all sites, the distance between the mean and median in the box plot remained unchanged in the scenarios when temporal resolution of sampling increased from 1 hour to 15 days. Accordingly, the intervals between the 25 and 75 % quartiles were not notably modified among the analysed temporal resolutions, which means that, visually, the frequency distributions were unchanged along the scenarios.
On the other hand, the main changes between the different temporal resolutions are the substantial reduction of outliers below and above the extreme quantiles (i.e. 10 and 90 %, respectively) at all sites. This behaviour indicates that the SM dynamics are no longer correctly represented when the temporal resolution reduces. The losses of maximum and minimum values, as the temporal resolution increases, portray the inaccurate assimilation of the SM dynamics at coarser temporal resolutions.  Figure 4 shows the SM variability by depicting the temporal means of time-invariant (M tn ) and absolute time-variant [μ t̂( |A tn |)] components at each site. Site 3 presented the highest M tn , whereas sites 1 and 4 showed the smallest ones; similar values (below 23 mm) were noticed at temporal resolutions varying from 1 hour to 4 days (Figure 4a). It is worth noting the increase in M tn at sites 2 and 3 at a temporal resolution of 7 days, with a value above 27 mm at site 3. In addition, site 2 showed an M tn between 22 and 24 mm, but displayed a critical increase (near 26 mm) at 7 days. The M tn values at all sites were closer to the spatial-temporal mean [μ n (M tn )] at a temporal resolution of 15 days, except for site 3. The increase of M tn at 7 days, at both sites 2 and 3, can be explained by losses of minimum SM data in areas where the soil is saturated for long periods (Figure 2). On the other hand, the observed SMC reduction at 15 days can be explained by losses     of maximum SMC data. These contrasted behaviours, from temporal resolutions of 4 to 15 days, compromise and mischaracterise the SM in the GEW when coarser temporal resolutions of monitoring are set at the sites 2 and 3.

Temporal variability analysis of time-variant and time-invariant components
The temporal anomalies were considered as absolute values (|A tn |) in figure 4b to circumvent compensations among positive and negative values. It was possible to observe A tn changes at different temporal resolutions, although they were less perceptible than those in temporal mean (M tn ) cases. However, both plots showed a similar behaviour, namely unusual high values at 7-day temporal resolution. This result stemmed from data losses caused by temporal resolution decreases in the time series, which had a stronger effect on the spatial mean of the SM temporal mean [μ n (M tn )] than on the spatial mean of the absolute temporal anomalies [μ n (|A tn |)]. The invariant term was then more sensitive than the variant one. Therefore, SM time-stability was slightly spatially affected by temporal anomalies (A tn ) in the GEW, since the temporal mean of the absolute temporal anomalies [μ t̂( |A tn |)] remained roughly unchanged at all sites.

Analyses of rank stability changing at different temporal resolutions
Rank stability was performed to analyse and compare the absolute and relative differences of the SM components at different temporal resolutions at all sites ( Figures 5, 6, and 7). However, only the results for the temporal resolutions at which the site ranking changed from 1-hour temporal resolution were displayed due to the large amount of data used in this analysis. The temporal resolutions of 4, 7, and 15 days were then plotted together with the 1-hour reference scenario because the temporal resolutions ranging from 6 h to 2 days remained unchanged when compared to the original scenario. Figure 5 shows the ranking stability plots of the calculated and absolute temporal means of absolute SM differences [μ t̂( ΔS tn ) and |μ t̂( ΔS tn )|, respectively] and of relative SM differences [μ t̂( δS tn ) and |μ t̂( δS tn )|, respectively]. At first, it is possible to notice that μ t̂( δS tn ) rankings remain the same at all analysed temporal resolutions. Additionally, sites 2 and 4 showed an SM temporal variability closer to the mean behaviour at absolute scale, because their μ t̂( ΔS tn ) values were closest to zero at almost all temporal resolutions. On the other hand, sites 3 and 1 were closer to the mean behaviour at relative scale, since the μ t̂( δS tn ) values were nearly zero. Likewise, the 4-days temporal resolution revealed a double swapping of rank positions for |μ t̂( δS tn )| between sites 1 and 3 and sites 2 and 4 (Figures 5m and 5n), remaining the same at 7-day and 15-day temporal resolutions (Figures 5o and 5p). Such consideration also took place for μ t̂( ΔS tn ) and |μ t̂( ΔS tn )| at 7 days, which alternated the rank positions between sites 1 and 4 at 1st and 2nd and 2nd and 3rd ranks, respectively (Figures 5c and 5g). Despite the change in rank positions between the two first non-reference temporal resolutions (from 4 to 7 days), the ranking of the sites for μ t̂( ΔS tn ) and |μ t̂( ΔS tn )| returned to the original setting at 15-day temporal resolution. These behaviours upheld the SM data losses at coarser temporal resolutions at sites 1, 2, and 4, but not at site 3. Therefore, the rank positions of most sites were sensitive to data losses entailed by reducing the temporal resolution, leading to ranking mischaracterisation. Figure 6 shows the rank stability plots of the temporal means of the absolute SM temporal means and the absolute anomaly differences [the μ t̂( |ΔM tn |) and μ t̂( |ΔA tn |), respectively]. Importantly, the μ t̂( |ΔM tn |) ranking had the same order than the |μ t̂( ΔS tn )| ranking (Figures 5e to 5h and 6a to 6d). Site 2 was the most representative in terms of the SM spatial mean within the monitoring network for all temporal resolutions, while site 3 was the least representative site (Figures 6a to 6d), even holding a similar soil type and prolonged recession curves. Furthermore, the μ t̂( |ΔM tn |) tended to decrease as the temporal resolution became coarser, except at the 7-day temporal resolution, when almost all sites (1, 2, and 4) presented an opposite tendency. Thus, the longer the temporal resolution of data collection, the smaller the μ t̂( |ΔM tn |) magnitudes, since the SM time series become more susceptible to anomalies in A tn . As such, the sites tend to decrease their amplitudes in Rev Bras Cienc Solo 2019;43:e0180236 different proportions, depending on the soil features, which may trigger an alteration in the ranking. The ranks of μ t̂( |ΔM tn |) and μ t̂( |ΔA tn |) swapped positions between sites 1 and 4 (Figures 6a and 6c) and between sites 1 and 2 (Figures 6f and 6g) at 7-day temporal resolution, respectively. This result corroborates with those results ascertained in the descriptive statistical analysis (see section 4.1). Therefore, the results suggest that the real SM processes are mischaracterised at a temporal resolution of 7 days, since the ranking of their temporal components changed in comparison to the reference scenario; the mentioned temporal resolutions show magnitudes higher than those in the reference scenario.   Figure 7 shows the scatter plots of the ranking of the absolute temporal means of absolute SM differences [|μ t̂( ΔS tn )|] against the rankings of μ t̂( |ΔM tn |) and μ t̂( |ΔA tn |). Sites 2 and 3 presented the lower and higher ranks for both relationships, respectively, without relevant variations (only site 2 in figure 7g) by changing the temporal resolutions. Conversely, sites 1 and 4 presented different ranks in the relationship of |μ t̂( ΔS tn )| against μ t̂( |ΔM tn |) and |μ t̂( ΔS tn )| against μ t̂( |ΔA tn |. It is highlighted that the ranking of the study sites has the same relationship between |μ t̂( ΔS tn )| and |μ t̂( ΔM tn )| for all analysed temporal resolutions, meaning that sites 1 and 4 swapped their ranks between the 2nd and 3rd positions in both metrics (Figure 7c). On the other hand, alternations in the ranks of |μ t̂( ΔS tn )| and μ t̂( |ΔA tn |) were observed at a temporal resolution of 7 days, where sites 1 and 2 swapped their 2nd and 1st positions in |μ t̂( ΔS tn )|, respectively, and sites 4 and 1 swapped their 2nd and 3rd positions in μ t̂( |ΔA tn |), respectively ( Figure 7g). Accordingly, the results presented in this analysis show that the invariant mean parcel mainly rules the SM variability.

Hierarchical cluster analysis
Hierarchical cluster analysis was carried out for all study sites by using the SM temporal components for all considered temporal resolutions. However, no relevant cluster reordering was found by changing the analysis for temporal resolutions longer than 1 hour, since a nearly stable and robust cluster association was found between SM components at all sites. Accordingly, the dendrogram for all study sites, using SM temporal components at 1-hour temporal resolution, is shown in figure 8. Overall, sites 1 and 4 had similar dendrograms, which is likely substantiated by the similarity of their soil type, land covers and topographic features (Figure 1). On the other hand, the dendrograms of sites 2 and 3 were similar, likely influenced by the Espodossolo (Podzol) flat areas and land covers (sparse Atlantic Forest and bush vegetation), which have a sandy texture in the first soil horizon (high permeability) and a restrictive layer nearby at a depth of ~2 m.   Rank: µ t (|ΔAtn|) (-) Rank: µ t (|ΔStn|) (-) Figure 7. Scatter plots of absolute temporal mean ranking of absolute soil moisture differences [|μ t̂ (ΔS tn )|] against the temporal mean ranking of absolute SM temporal mean differences μ t̂ (|ΔM tn |) (a, b, c, and d) and temporal mean ranking of absolute temporal anomaly differences μ t̂ (|ΔA tn |) (e, f, g, and h). Based on the hierarchical clustering, it can be inferred that S tn and A tn had fast links at all sites. The relative D link /D max for these two variables ranged from 1 to 3 %, mainly due to the influence of the dynamic parcel on the SM behaviour. This linkage reinforces that temporal anomaly is the primary SMC component reflecting the temporal variability patterns, irrespective of land cover, soil type and slope (Figure 1). A similar linkage was found between variables ΔS tn and ΔA tn , with D link /D max varying from 2 to 5 %, followed by δS tn values ranging from 11 to 29 %. This behaviour denotes the reasonable influence of the dynamic parcel ample oscillation on the magnitude of SM temporal variability.
On the other hand, M tn and ΔM tn seemed to link belatedly to S tn and A tn at sites 1 and 4, with D link /D max varying from 52 to 71 %, respectively; however, they were completely separated at sites 2 and 3 (D link /D max equals 100 %). This result corroborates with previous findings of this study, showing the negligible SM time-variant effects over the overall time-invariant aspect of SM dynamics at the downstream sites of GEW compared to the upstream ones.

DISCUSSION
Soil moisture (SM) temporal dynamics at downstream behaved differently from the upstream of the border between two tropical soil types, similar to the findings reported in small-scale watersheds by Dias et al. (2015) and Hu et al. (2017), but not to the results of other studies (Qiu et al., 2001;Zhou et al., 2015). Ranking rearrangements at 7 days were incited by some weekly rainfall peaks, associated with prolonged soil saturation events during the rainy season (mainly at sites 2 and 3), which led to losses of minimum values (Figure 6), whose effects have also been reported by Zhang and Shao (2017). Moreover, the slight decrease in μ t̂( |ΔM tn |) at site 3 was likely imposed by non-processing the high SMC peaks at coarser temporal resolutions, since high values with slow recession curves are more frequent during the wet season at this site (Figure 2), which has also been found by Molina et al. (2014).
The reason the SM time series at sites 1 and 4 show smaller amplitudes despite the sandy soil texture is that these sites feature either a moderately sloped topography or the absence of a shallow restrictive layer, characteristic of Argissolo Vermelho-Amarelo (Haplic Acrisol). These properties increase the water residence time in the Argissolo (Acrisol), yielding a deeper groundwater recharge, similar to Amazon Espodossolo (Podzol)-to-Argissolo (Acrisol) soil transition (Nascimento et al., 2008). The sandy texture in Argissolo Vermelho-Amarelo (Haplic Acrisol) is in agreement with the GEW hydrodynamic characterisation (Santos et al., 2012), but differs from the overall Argissolo (Acrisol) texture elsewhere (Montenegro and Ragab, 2012;Ottoni et al., 2018). This combination of soil properties demonstrates the particular condition of the study area.
Around the border between the two soil types, site 2 is a representative spot in the middle of the GEW, mainly because is better represented both the SM temporal variability ( Figure 5) and its invariant and dynamic parcels at the reference scenario ( Figure 6). The high peaks and the quick recession curves stem from the capillarity rise effect due to the shallow groundwater table, which comes close to saturation during the rainy season (Figure 2), as previously described by Montenegro et al. (2018) in Cambissolo Háplico (Sodic Cambisol). Still, at upstream of the border between the two soil types, the waterlogging lasted longer at site 3 because of its proximity to the watershed boundary, where the groundwater is shallower. These behaviours typically occur in thin, flat soil packages associated with restrictive layers (duripan in this study), because they hold a high water permanence subjected to sub-surface percolation processes .
Although sites 2 and 3 showed the highest maximum SM peaks along the rainy season, they also showed the lowest SM values during the dry season. The slightly higher SM in the dry season at downstream is likely related to the presence of mulch (Souza et al.,

CONCLUSIONS
The present study demonstrated that varying the soil moisture (SM) monitoring time-step at some sites and ranking the mean statistical metrics of their SM time-variant and time-invariant parcels may point out changes on the stability of SM temporal dynamics. It is concluded that around the border between tropical soil types, the distinct behaviours of time-variant and time-invariant components of SM time series reflect the different combinations of their soil properties, namely topographic slope, depth to the restrictive layer, and soil texture. Moreover, the steady soil properties over time, such as slope and soil type, regulate the SM temporal mean, whereas the variable properties in time, such as rainfall events and groundwater table, control the SM temporal dynamics. The rearrangements in site rankings mainly stem from differences in the time-invariant parcel, which govern SM temporal variability over the time-variant parcel, either at high or low temporal resolutions. Lastly, losses of SM monitoring resolution mainly affect the SM time-variant parcel, hinting at the potential mischaracterisation of the SM time-stability in tropical soil types for time-steps longer than 2 days.    Figure 5. Ranking stability plots of temporal means of absolute Soil Moisture (SM) differences [μ t̂ (ΔS tn )] (a, b, c, and d), absolute temporal means of absolute SM differences [|μ t̂ (ΔS tn )|] (e, f, g), and temporal means of relative SM differences [μ t̂ (δS tn )] and absolute temporal means of relative SM differences [|μ t̂ (δS tn )|] (i, j, k, and l).