A statistical comparison of spatio‐temporal surface moisture patterns beneath a semi‐natural grassland and permanent pasture: From drought to saturation

Some 60% of the agricultural land in the UK is grassland. This is mostly located in the wetter uplands of the west and north, with the majority intensively managed as permanent pasture. Despite its extent, there is a lack of knowledge regarding how agricultural practices have altered the hydrological behaviour of the underlying soils relative to the adjacent moorland covered by semi‐natural grassland. Near‐surface soil moisture content is an expression of the changes that have taken place and is critical in the generation of flood‐producing overland flows. This study aims to develop a pioneering paired‐plot approach, producing 1,536 moisture measurements at each of the monitoring dates throughout the studied year, that were subsequently analysed by a comparison of frequency distributions, visual cum geostatistical investigation of spatial patterns and mixed‐effects regression modelling. The analysis demonstrated that the practices taking place in the pasture (ploughing, re‐seeding and drainage) reduced the natural diversity in moisture patterns. Compared to adjacent moorland, the topsoil dried much faster in spring with the effects requiring offset with moisture from slurry applications in summer. With the onset of autumn rains, these applications then made the topsoil wetter than the moorland, heightening the likelihood of flood‐producing overland flow. During the sampling within one such storm event, the adjacent moorland was almost as wet as the pasture with both visibly generating overland flow. These contrasts in soil moisture were statistically significant throughout. Further, they highlight the need to scale‐up the monitoring with numerous plot pairs to see if the observed highly dynamic, contrasting behaviour is present at the landscape scale. Such research is fundamental to designing appropriate agricultural interventions to deliver sustainable sward production for livestock or methods of mitigating overland‐flow incidence that would otherwise heighten flood risk or threaten water quality in rivers.


| INTRODUCTION
Grassland accounts for 60% of the total UK agricultural area, which is proportioned almost equally at 55% as agriculturally improved permanent pasture and 45% rough grazing on semi-natural grasslands (DEFRA, 2019). Both permanent pasture and semi-natural grassland (often referred to as open-moorland or 'unimproved' pasture) encompasses a large percentage of the UK uplands, providing sustenance to grazing livestock alongside other ecosystem services (Bengtsson et al., 2019;Hayhow et al., 2019;Lamarque et al., 2011;Morse, 2019). Historically, semi-natural grassland was converted into permanent pasture during the eighteenth and nineteenth century to increase agricultural output (Kain, Chapman, & Oliver, 2004;Whyte, 2006). Gilman (2002), O'Connell et al. (2004), Holden et al. (2007 and Wheater et al. (2008) all note the lack of research into the hydrological functioning of semi-natural grassland, with Gilman (2002) directly stating that 'there is little or no experimental evidence to support theoretical studies' relating to the effects of seminatural grasslands on flood risk. Consequentially, it remains unknown how converting upland semi-natural grassland into permanent pasture has altered soil moisture regimes that affect flood generation processes and drought resilience.
Very few upland UK studies have compared semi-natural grasslands to permanent pasture, with research operating at coarse-scale resolution without conducting paired-plot analysis, thus, observations and knowledge of hydrological processes at the plot-hillslope scale is lacking. Orr and Carling (2006) compared catchment scale flood risk within North-West England, commenting that transitioning from Heather (Calluna vulgaris) or scrub vegetation to drained pasture could increase downstream flood risk. Marshall et al. (2006) and Wheater et al. (2008) similarly concluded, through hydrograph assessment, that a semi-natural grassland in mid-Wales, UK (Pontbren experimental site), had a damped flood response compared to improved pasture. Ockenden and Chappell (2008) noted that a measured plot of seminatural grassland was significantly drier than a nearby improved pasture in the River Eden catchment (Cumbria, UK). McIntyre and Marshall (2010), at Pontbren, noted that semi-natural grassland tended to have a less flashy flood response than improved pasture. Gilman (2002) is the only UK study to compare permanent pasture to semi-natural grassland as a primary research objective. The study concluded that pasture reversion could reduce River Severn peak flows by 0.5-2%, and smaller channel peak flows by 2-4%, although acknowledged the lack of supporting studies with which to justify model values used to simulate changes. These studies emphasise a considerable research gap, justifying the need for a hydrological comparison of permanent pasture and semi-natural grassland in an upland UK landscape (Wheater et al., 2008). Indeed, there is a global dearth of studies relating to how livestock production alters the hydrological functioning of natural soils (Magliano et al., 2019).
A significant component of the catchment water budget is the soil volumetric wetness (θ V ), which is the total volume of water present between soil particles divided by the total undisturbed soil volume.
The aim of this study is to compare the spatio-temporal dynamics of surface soil volumetric wetness in an area of semi-natural grassland with an adjacent area that has been converted and managed as permanent pasture in the UK uplands. Both the reference and converted plots are adjacent to minimise natural differences. The methodological development aspect of the research aims to quantify the spatial variability of θ V at the plot scale, which demands intense measurements.
The research specifically measures each plot temporally, rather than the replication of plot pairs in the landscape, which is beyond the scope of this study. The plot comparison is conducted over a 6-month period (including drought and fully-saturated conditions), to assess non-stationarity in the differences. A high-resolution (1 m 2 ) volumetric wetness grid (1,536 m 2 ) was needed to capture fine-scale spatiotemporal soil moisture variability, which is then compared with localised factors such as land-use, vegetation, season, and elevation, to assess their impact.
Thus, the detailed research objectives are: 1. To develop a statistically robust methodology for the quantification of soil moisture differences between an example 768m 2 area of semi-natural grassland and an adjacent area of the same size managed as permanent pasture.
2. To statistically contrast the volumetric wetness probability density functions between a permanent pasture and a semi-natural grassland, to quantify soil moisture differences.
3. To compare geostatistically the spatial structure of soil volumetric wetness between a permanent pasture and a semi-natural grassland, to assess spatially dependent soil moisture patterns. 4. To determine which factors significantly influence volumetric wetness in the contrasting, adjacent land-uses, to highlight any potential predictors of volumetric wetness at this particular locality.

| Study site
Measurements were taken within a permanent pasture and a bordering semi-natural grassland located within the Lowther catchment, 3 km northwest of Shap (Cumbria, UK), between April 2018 and May 2019. The pasture (centre 54 32 0 26" N, 2 43 0 44" W) and seminatural grassland (centre 54 32 0 26" N, 2 43 0 42" W) are immediately adjacent and are separated by a 1.3 m drystone wall (Figure 1). The drystone wall was likely raised between 1838 and 1855 based on surrounding Enclosure Acts (Kain et al., 2004;Whyte, 2006). Both plots are mapped as Brickfield Soil Association . This equates to an FAO Eutric Stagnosol, or an Aquic soil within several USDA soil orders (USDA, 1999;WRB, 2015). Eutric Stagnosols are widespread throughout the UK uplands, and are highly susceptible to saturation, poor drainage, and overland flow . The study site soils are till derived and slowly permeable, which overlay Tarn Moor Formation mudstone of the Buttermere and Bitter Beck Formations within the Skiddaw Group (Cooper et al., 1995;Stone, 2007).
The local climate from the Shap weather station (54 30 0 49" N, 2 40 0 40" W: 301 masl: Figure 1) is wet temperate, with a mean winter temperature of 4.1 C, a mean summer temperature of 11.5 C, and an annual rainfall average of 1,779 mm (Met F I G U R E 1 The experimental site within a UK, Cumbrian, and local area context. The permanent pasture (PP) and semi-natural grassland (SNG) sites are highlighted in green stripes and pink crosshatch, respectively. Shap weather station, alongside the downstream river gauging station (Eamont Bridge), is shown. Historically, the pasture was semi-natural grassland until being improved during the Inclosure (Enclosure) Acts of the early-mid 19th century, with the wall likely erected between 1838-1855. The site is within a headwater where downstream settlements such as Penrith, Eamont Bridge and Carlisle suffer from flooding Office, 2020). Daily precipitation data alongside downstream River Lowther discharge (gauged at Eamont Bridge; 54 38 0 60" N, 2 44 0 15" W: 119 masl: Figure 1) during the study is given in Figure 2. An Antecedent Precipitation Index (API) for the study site ( Figure 2) was calculated according to Equation ((1)): where API is the antecedent precipitation index, R is the daily precipitation total and κ is an empirical decay factor below 1. A κ value of 0

| Experimental design
A paired-plot experimental design was adopted as PP and SNG are immediately adjacent Eutric Stagnosols with similar slopes (4-4.5%).
Both PP and SNG have virtually identical distributions of topographic wetness ( Figure 3). Both plots were covered by semi-natural grassland until PP was enclosed, likely during the early-mid 19th century (Kain et al., 2004). This experimental design therefore suggests observed differences are due to land conversion and subsequent management as opposed to inherent site dissimilarity. The study site location was appropriate due to PP and SNG belonging to the most common upland soil type in England, with both sites following typical regional pastoral/moorland agricultural practice. Supporting precipitation and discharge information was available to infer site conditions prior to and between sampling and to aid interpretation of results. Given that frontal rainfall is the dominant precipitation mechanism in the UK, The daily precipitation data taken from Shap weather station throughout the experiment, alongside the Antecedent Precipitation Index (API) and the mean daily downstream flow at Eamont Bridge (see Figure 1). The sampling dates and API for 29th May (407) wave. The outgoing and standing wave ratio is dependent upon the dielectric constant of the soil surrounding the waveguides, which is largely controlled by θ V (see Gaskin & Miller, 1996). Simplified TDR was the selected experimental method due to it being a rapid and repeatable in-field technique (see Gaskin & Miller, 1996).
Following Whalley (1993), the moisture probe reported in-field measurements in millivolts (mV), which were converted to θ V postmeasurement via the calibration equation (Equation (2)): θ V = 1:07 + 6:4mV − 6:4mV 2 + 4:7mV 3 where α 0 and α 1 are soil coefficients and were taken as −1.6 and + 8.4, respectively, due to the experiment primarily involving mineral soils (Whalley, 1993). The moisture probe is accurate to ± 2% θ V and averages θ V over the full length of the waveguides, primarily around the central waveguide (Gaskin & Miller, 1996;Whalley, 1993). The same moisture probe was used for all measurements to account for any unknown instrument bias. Gaskin and Miller (1996) and Miller, Gaskin, and Anderson (1997) give detailed information regarding the design, operation, calibration and uncertainty of moisture-probe measurements.

| Reference topsoil physio-chemical properties which may influence volumetric wetness
Several large-scale studies have shown that soil physio-chemical properties can substantially influence soil θ V (e.g., Pan, Boyles, White, & Heitman, 2012). Soil texture (Wallace & Chappell, 2019), porosity (Beven & Germann, 1982), acidity (Holland et al., 2018), bulk density (Drewry, Littlejohn, & Paton, 2000), penetration resistance (Wallace & Chappell, 2019) and organic matter content (Beven & Germann, 1982) can all affect soil structural stability and functioning, and therefore permeability and water retention. To determine such properties, topsoil was extracted from the surface 10 cm. Soil samples were taken using 221 cm 3 bulk density tins across 16 randomly selected locations throughout PP and SNG ( Figure 4). Random sampling was chosen as it eliminated sampling bias. Four soil samples per location were taken, with three undergoing an initial 48-hour air dry. The first air-dried sample was used to measure soil pH. The second air-dried sample was oven dried at 105 C for 24 hr for dry bulk-density calculation, and then underwent a 6-hr 550 C loss-on-ignition test to calculate organic matter (OM). The third air-dried sample underwent particle size analysis. Particle size analysis involved sieving oven dried soil through a 2000-μm sieve, before mixing the sample with 1% sodium polymetaphosphate for 24 hours to separate aggregates. The soil then underwent hydrogen peroxide treatment to remove organic material.
Finally, samples underwent manual aggregate breaking and high-power sonication for 5 min before laser diffraction (Beckman Coulter, LS-13-320). The final sample was gradually submerged for 48 hr with de-ionised water and was then measured with the moisture probe to determine soil porosity (i.e., maximum volumetric wetness) to remain consistent with field measurements. Soil penetration resistance was measured in situ adjacent to soil sampling locations, using an SC900 Field Scout (Spectrum Technologies) penetrometer using a 12.8 mm diameter cone. The device measures soil penetration resistance via an internal load cell and uses an ultrasonic depth sensor to record depth in 2.5 cm steps for up to 7.5 cm. to allow θ V to be compared against floristic composition. Vegetation potentially explains a significant amount of θ V variance within areas of (semi-)natural vegetation (Chappell & Ternan, 1992;Meijles et al., 2003). The genera/species that encompassed the majority of the above-surface biomass in each square metre was recorded, even if several were present. Each square metre was centred around a TDR

| Volumetric wetness spatial structure (objective II)
The GLOBEC geostatistical package in MATLAB (Chu, 2017) was used to generate empirical semi-variograms from θ V observations, to assess PP and SNG spatial structure. Semi-variogram models were derived using the GLOBEC least-square fit function. A range of models were fitted to the empirical semi-variogram data, including exponential (Equation (3)), Gaussian (Equation (4)) and spherical (Equations (5.1) and (5.2)) models: where γ h is semi-variance at each lag distance, P 0 is the partial sill, h is the lag distance, L is the length scale and γ 0 is the nugget effect. A residual sum of squares gave a goodness of fit for each semivariogram model.
During topsoil sampling (Table 2), PP was significantly drier than SNG (p ≤ .001), with a median θ V of 39% as opposed to 47%. The pasture also had significantly higher soil penetration resistance (p ≤ .006-.041) at all recorded depths (0, 2.5, 5 and 7.5 cm), partly due to the drier soil (Wallace & Chappell, 2019). Bulk density was significantly higher (p ≤ .028) within the PP plot compared to the SNG plot, with porosity significantly lower (p ≤ .022), possibly indicating that vegetation differences may have some influence upon soil properties (Macleod et al., 2013), or that agricultural practices had compacted the pasture and potentially reduced the infiltration capacity (Drewry et al., 2000;Gilman, 2002;Pan et al., 2012).
This may be because of slurry additions to the pasture that are maintaining the naturally high levels of OM seen in the soils beneath semi-natural grassland. Organic matter content was, however, considerably more variable within the SNG plot, likely due to the presence of localised carbon-rich 'rush flushes' within the Eutric Stagnosol (Chappell & Ternan, 1992). Indeed, sample point 'N' in SNG (Figure 4) contained more than 2.5 times the OM of the most organic PP sample. The PP plot was significantly less acidic than the SNG plot (p ≤ .001), probably due to liming (Holland et al., 2018).

| Topography and elevation survey
The detailed topographic survey ( Figure 6) showed that PP was marginally higher than SNG, with an arithmetic mean of 222.5 masl as

| Vegetation survey
The taxonomic survey (Figure 7) reveals a small number of dominant genera/species within each square metre. The PP plot was almost entirely dominated by ryegrass, encompassing 94.7% of the sampling grid, with pockets of common rush covering only 4.4%. Stinging nettle (Urtica dioica) and broad-leafed dock (Rumex obtusfolious) were present against the drystone wall, at 0.7 and 0.3% of the area, respectively. The pasture contained significant clover, buttercup (Ranunculus spp.) and ox-eye daisy (Leucanthemum vulgare) populations, likely from re-seeding mixtures, although these never dominated a grid cell.
Vegetation within SNG (Figure 7) was predominantly a mixture of the Pooideae subfamily of grass species (65.4%), primarily consisting of common bent (Agrostis capillaris), creeping bent (Agrostis stolonifera), mat-grass (Nardus stricta) and ryegrass. This grass mixture is hereby referred to as 'moorland grass'. Common rush was also very common at 33.6%. Stinging nettle and broad-leafed dock occurred at F I G U R E 5 The soil particle-size analysis for the Permanent Pasture (PP) and the Semi-Natural Grassland (SNG) 0.7 and 0.3% incidence, respectively. A large number of additional vegetation species were recorded within SNG, particularly plume thistles (Cirsium spp.), and Sphagnum mosses (Sphagnum spp.), with no grid containing fewer than three species.
Ryegrass dominance within PP is due to pasture re-seeding to maintain sward levels and prevent reversion (Gilman, 2002), with the significant clover and ox-eye daisy population further supporting reseeding. Common rush prevalence in both land-uses is likely caused The elevation profile at the study site, within the Permanent Pasture (PP: top) and Semi-Natural Grassland (SNG: bottom). The wall separating the plots is approximately along a NNW-SSE axis. Note the depression at the boundary within SNG at approximately 28 m distance parallel to the wall, which extends further into SNG. It is likely this local depression will remain wetter than the surrounding areas throughout the experiment (see Figure 3) The dominant vegetation within each square metre, with the Permanent Pasture (PP: top) and Semi-Natural Grassland (SNG: bottom) highlighted. Note that SNG contains substantially more common rush, and that the moorland grass is an admixture of several different grass species. The stinging-nettle and broad-leafed dock within PP border the wall and are amassed around a small tree stump by locally poor drainage in a high rainfall environment, both highly suited to Juncus spp. proliferation (AHDB, 2013;McCorry & Renou, 2003). Gilman (2002) notes that Juncus spp. will often be the first moorland species to colonise ryegrass/clover swards in high rainfall upland environments of the UK. Nearby upland catchment studies have also noted considerable Juncus spp. and Nardus spp. compositions, both being typical of upland soils (Gilman, 2002;Orr & Carling, 2006).
To reemphasize, minimal research has compared the differences between moorland and pasture vegetation that may be correlated with changes in hydrological properties. Semi-natural grassland conversion likely reduced vegetation height, biomass and root depth, thereby reducing rainfall input due to wet-canopy evaporation, as well as soil porosity (Gilman, 2002;Orr & Carling, 2006;Sansom, 1999). Sansom (1999) and Gilman (2002) postulate that moorland conversion reduces infiltration rates, hydraulic conductivity, surface roughness and evapotranspiration, ultimately causing increased overland flow and elevated flood risk.

| Volumetric wetness probability distributions (objective I)
Tables 3 and 4 highlight that no PP or SNG probability distribution satisfied normality (KS or AD tests), justifying the use of nonparametric statistical tests. May and August distributions had predominantly weak positive skews, whilst October and November had more extreme negative skews (Figure 8). Distributions were principally leptokurtic (displaying excess kurtosis), especially November (Tables 3 & 4). Box-Cox transformations could only normalise (AD tests only) May and August PP distributions, further justifying the non-parametric approach.

| May dataset
During the May sampling date (Figure 9), PP was significantly drier than SNG (p ≤ .001), with a median θ V of 27.6% as opposed

T A B L E 3
The Kolmogrov-Smirnov (KS) and Anderson-Darling (AD) statistical distribution tests applied to the Permanent Pasture (PP) and Semi-Natural Grassland (SNG) volumetric-wetness probability distributions Note: Note that most tests are extremely significant, indicating that they significantly differ from the Gaussian distribution. The statistical tests are paired with excess kurtosis and skewness values to infer why distributions may violate normality. Box-Cox transformed distributions using the maximum log-likelihood function are shown adjacent to the raw data to highlight the extent of non-normality, with further supporting kurtosis and skewness values. *Significant at the 0.05 probability level. **Significant at the 0.01 probability level. ***Significant at the 0.001 probability level.

T A B L E 4
The summary statistics, including arithmetic mean (x), median (x), and coefficient of variation (CV), for the volumetric wetness measurements taken at each sampling date within the Permanent Pasture (PP) and Semi-Natural Grassland (SNG) to 42.7% (Table 4: Figure 8). Volumetric wetness differences could be because of higher evapotranspiration within PP, due to the rapidly growing, dense ryegrass sward (Cox, Parr, & Plant, 1988;Hall, 1987). Evapotranspiration is generally assumed to be greater from semi-natural grassland compared to pasture; although this relationship is dependent on vegetation growth and is primarily based off studies involving heather as opposed to 'rush pasture' (Gilman, 2002;Hall & Harding, 1993;Miranda, Jarvis, & Grace, 1984;Orr & Carling, 2006).
During May, the SNG plot had several unvegetated soil patches, the moorland grass was heavily grazed and the common rush was withered with minimal foliage, all implying low transpiration rates. Furthermore, the PP plot could additionally contain fewer pockets of impermeable soil as local agricultural practices encourage drainage, reducing θ V (Wallace & Chappell, 2019).
Removing θ V variability within a pasture is a central objective of ploughing prior to re-seeding, in order to generate an even grass sward (Schulte et al., 2012). This pioneering study has shown that the PP plot did indeed contain significantly less variation in θ V (p ≤ .001) than observed in the SNG plot (Table 4: Figure 8). If the ecological status and functioning of permanent pastures were to be restored to behave more like semi-natural grassland, then the diversity in moisture patterns would need to be re-introduced. As this first sampling date (spring 2018) shows the pasture to be drier than the semi-natural grassland, if representative, this may suggest that permanent pastures dry faster and thus are more sensitive to water stress with the onset of droughts, a potential concern for livestock production.

| August dataset
Between May and August 2018, the semi-natural grassland saw median θ V fall from 42.7 to 23.4% (Table 4, Figures 8 & 10). At the 2nd August 2018 sampling date, the API was close to the lowest value for the whole of 2018 (Figure 2), indicating that the sampling programme had observed soil near its driest state in 2018. The high degree of drying was due to relatively high levels of solar radiation over the summer months and moderate rainfall since the previous measurement (only 192 mm in 64 days), with the 40 days prior to sampling recording 56% of the long term average rainfall for this period at this locality (Met Office, 2020). In some contrast, the median θ V in the PP plot was maintained over the same period, increasing slightly from 27.6 to 30.4% (Table 4, Figure 8). As a result, the SNG plot became significantly drier than the PP plot (p ≤ .001). As before, the SNG plot contained significantly more variance than the pasture (p ≤ .001).
The additional drying effects of higher radiation and lower rainfall were more than offset by artificial moisture additions in the form of slurry to the PP field. This indicates that while the PP plot initially F I G U R E 8 The kernel-generated probability density functions for the volumetric wetness during each sampling date. Each distribution was generated according to 768 samples based on the respective land-use. Note that these are statistically tested for normality in Table 3, and the central tendency and variation is statistically compared between land-uses in Table 4 dried faster than the SNG plot, agricultural interventions could offset these effects. Slurry additions to pastures do have a negative impact on the water quality of adjacent streams however (Hunter, Perkins, Tranter, & Gunn, 1999). Consequently, if such additions were not permitted, then the permanent pasture would lose its artificial moisture input during a drought, and from the May results, could be in a drier state than the semi-natural grassland when the drought is most severe. Withholding slurry during these periods would therefore likely cause substantial sward damage (Schulte et al., 2012).

| October dataset
Over the 81 days between the 2nd August and 23rd October 2018 sampling dates, 504 mm of rain was recorded ( Figure 2). As a result, the SNG plot became much wetter, increasing to a median θ V of 46.6% (Table 4, Figures 8 & 11). The PP plot became wetter still; increasing to 53.6% (Table 4, Figure 8) and remaining statistically wetter than the SNG plot (p ≤ .001). Interestingly, θ V within both plots increased by the same amount (+23.2%). Identically to previous sampling dates, the SNG plot contained significantly higher variance (p ≤ .001).
The October θ V data show that pastures with summer slurry additions can be wetter than semi-natural grasslands at the onset of autumn rains. Indeed, the median pasture θ V is only 4.6% below the median porosity, suggesting that most of the pasture is near saturation and could quickly saturate during storm events. The semi-natural grassland θ V is 12.9% below median porosity, suggesting some remaining storage capacity before SOF generation. At moisture plots 20 km to the East of those in this study, Ockenden and Chappell (2008) also observed that their single permanent pasture plot was wetter than semi-natural grassland plots during autumnal monitoring.

| November dataset
November sampling (Figure 12 Figure 8). Variance remained significantly higher in SNG, however (p ≤ .001: Table 4). Ockenden and Chappell (2008), working at the sites previously mentioned, similarly observed larger θ V variation within semi-natural grasslands compared to pastures for monitoring dates including the winter.

F I G U R E 9
The volumetric wetness grid taken on the 29th May 2018. Note that the Permanent Pasture (PP) is at the top of the figure, and the Semi-Natural Grassland (SNG) is at the bottom, with the wall shown to separate the land-uses. The permanent pasture was significantly drier than the semi-natural grassland, and contained significantly less variation. Linear features draining from the south-west to north-east according to the 'regional' topographic highpoint are also evident, primarily within SNG Between October and November sampling dates, 265 mm of precipitation fell in 36 days, and this was reflected in a very high API on the sampling date ( Figure 2). The strong negative skew within both frequency distributions suggests that moisture content at most places in both plots was approaching the upper limit of topsoil wetness, that is, the porosity (Tables 2 & 4: Figure 8: Western et al., 2002). These findings suggest that during large storm events, even semi-natural grasslands may generate SOF and so heighten local flood-risk. Thus attempting to re-establish semi-natural grasslands and associated soils in areas of permanent pasture may not necessarily reduce the incidence of SOF as part of so-called Natural Flood-risk Management.

| Volumetric wetness spatial structure (objective II)
The geostatistical analysis shows that the spatial structure of θ V within the SNG plot remained similar (i.e., relatively stationary) from May to November 2018 and is described well by exponential/spherical models (Figure 13; Table 5). Meijles et al. (2003) identically found semi-natural grassland at Dartmoor, UK, to have exponential or spherical semi-variogram models.
The spatial structure of θ V within PP was similar to that of the SNG plot during the relatively dry conditions of May. Slurry additions and rainfall gradually shifted the spatial structure from exponential to a Gaussian relationship, whereby the autocorrelation continued beyond the size of the experimental plot ( Figure 13). Selected models suitably fit the pasture data, although the October semi-variogram has noticeable residuals at large lags, probably because of the transitioning spatial structure.
The sill is the point at which the semi-variance plateaus within a model (i.e., the semi-variance as lag distance approaches infinity).
Most semi-variogram models ( Figure 13; Table 5) have a sill marginally above 1, with October PP having a slightly higher sill of 1.31, and November PP (Gaussian model) having a sill at 6.91. The elevated sills outline higher spatial variance of two distantly separated points as the pasture saturated, which was unobserved within SNG (Grayson & Blöschl, 2000).
The effective range is the distance from zero lag to the onset of the sill (95% in exponential models, 100% in spherical and Gaussian models) and can be interpreted as correlation length (i.e., the point beyond which there is no spatial autocorrelation).
The effective range within SNG remained essentially stationary throughout the study, suggesting θ V spatial autocorrelation is independent of the level of saturation. With increased saturation, PP contained considerably larger effective ranges than SNG ( Figure 13: Table 5). Agricultural interventions within PP likely homogenised soil variation and facilitated moisture redistribution.
As soils saturated, the lack of heterogeneity exerts a greater control on soil moisture redistribution at decimetre scales rather than The wall is shown to separate the land-uses. The pasture was significantly wetter than the semi-natural grassland at the time of sampling, and contained significantly less variation. Linear features are still clearly observable within the semi-natural grassland, although these are slightly masked in the permanent pasture due to the level of saturation F I G U R E 1 2 The volumetric wetness grid taken on the 29th November 2018 during Storm Diana. Note that the Permanent Pasture (PP) is at the top of the figure, and that the Semi-Natural Grassland (SNG) is at the bottom. The wall is shown to separate the land-uses. Both land-uses have statistically similar medians during these extremely saturated conditions, although SNG remained significantly more varied. Linear features are weakly observable within the semi-natural grassland, even though most of the land-use is at saturation the metre scales seen with the natural soils under the seminatural grassland, and thus, amplifies spatial autocorrelation. This was also seen within the decreasing coefficient of variation as PP saturated (Table 4). Ockenden and Chappell (2008) similarly found shorter correlation lengths for semi-natural grassland when compared with those of a single pasture plot. Meijles et al. (2003) found correlation length in semi-natural grassland to vary with saturation, an unobserved process in this study.
The nugget variance is the model semi-variance at zero lag and is generally interpreted as a combination of sampling/instrument error and spatial variation below the minimum sample spacing (i.e., <1 m variation). Within both PP and SNG plots, the nugget variance reached approximately half that of the sill variance ( Figure 13). This would indicate that there is significant variation in θ V at distances shorter than the 1 m sampling grid. This suggests that future studies that are able to collect more than the 1,536 (i.e., 768 × 2) values of θ V across a paired-plot on a sampling day should do so over an even finer sampling resolution (e.g., 10 cm grid). This would confirm whether deterministic spatial structure is present at sub-metre scales or whether other factors such as instrument-related uncertainty in θ V measurements are responsible. A simulated semi-variogram with 2% θ V error (uniformly distributed) for a plot-scale grid gave a nugget F I G U R E 1 3 The GLOBEC generated empirical semi-variograms for all sampling dates. All models are fitted with the least-squares fit function within GLOBEC, using the 768 samples in each land-use. Note that empirical semi-variograms become progressively dissimilar as the experiment proceeded, and that the full length scale lag distance is 48 m The model parameters from GLOBEC generated empirical semi-variogram models for each sampling date, alongside the Antecedent Precipitation Index (API) for each of the respective dates Note: The actual range in metres is given below the effective range. Note that the sill is the nugget plus the partial sill.
a Note that the model fit is reduced at larger lag distances and, therefore, that the true model effective range is, therefore, highly uncertain.
variance of approximately 0.05, suggesting instrument-related error is minimal (see Figure S1 and Appendix A). The slightly higher nugget variance in PP compared to SNG suggests increased fine-scale θ V variation within the pasture, further implying decimetre-scale moisture redistribution.

| Predictor variables of volumetric wetness (objective III)
The final objective of this study is to determine for this particular plot pair, the relative strength of the relationships between soil moisture content and the potential predictors of land-use, elevation, vegetation and season. This was assessed via correlation coefficients and mixedeffects regression modelling. Table 6 shows the correlation coefficients (r) of θ V and elevation for both individual and combined PP and SNG plots ( Figure 6). Weak correlations between topsoil moisture and elevation for the two plots suggest that elevation is not acting as a dominant control on θ V (see Figure 3). Combining the weak relationships with the limited topographic range justifies the use of elevation as a random effect in the linear mixed-effects regression model.
Beneath the complex vegetation communities of the semi-natural grassland, differences in θ V between common rush and grass species were not apparent (Table 7). This may be because of weak vegetation differentiation within the SNG plot, with many sampling grids containing both rush and grass species. Other studies have more successfully differentiated semi-natural grassland vegetation species, with Meijles et al. (2015) outlining that moorland grasses saturate faster than heather or bracken (Pteridium aquilinum). Conversely to the SNG plot, soil beneath common rushes in PP was wetter than ryegrass during August and October sampling dates, although similar in May and The correlations of volumetric wetness with elevation for both the Permanent Pasture (PP) and Semi-Natural Grassland (SNG), as well as when combined, expressed in terms of the correlation coefficient (r) Elevation and soil volumetric wetness correlation coefficient (r)

T A B L E 7
The arithmetic mean (x) and median (x) volumetric wetness for ryegrass and common rush within the Permanent Pasture (PP) and Semi-Natural Grassland (SNG) for each sampling date November (Table 7). While the mechanism is unclear, the dense root network beneath common rush apparently retains more moisture (from drainage or transpiration) following slurry or rainfall compared to soil beneath ryegrass.
As a measure of the relative importance of temporal changes (i.e., across the four sampling dates), spatial differences due to landuse (i.e., semi-natural grassland versus agriculturally improved permanent pasture) and vegetation (i.e., common rush versus grass species); linear mixed-effects regression modelling was undertaken and the results presented in The key findings were: • The contrast in soil moisture patterns between the paired plots changed markedly throughout the monitoring period, as did the interactions between the potential controlling variables. During spring sampling (29th May 2018), the pasture was significantly drier than the semi-natural grassland, making the vegetation more sensitive to water stress. With the reduced rainfall and higher transpiration of summer, the moisture content of the semi-natural grassland plot reduced to only 23%. In some contrast, moisture added in the form of cattle slurry maintained topsoil moisture at 30% in the pasture, underlining an overlooked agronomic benefit of slurry. As these slurry additions have consequences for water quality, a desire to restore wildlife habitats could see this practice barred. As the pasture was significantly drier than the semi-natural T A B L E 9 The regression model output giving variance (σ 2 ), slope effects of elevation: month (τ 00 Elev:Month ) and elevation: vegetation (τ 00 Elev:Vegetation ), the intra-class correlation (ICC), the number of elevation values (N Elev ), the total number of observations (N Obs ), marginal and condition R 2 values, the Akaike Information Criteria (AIC), and the Bayesian Information Criteria (BIC) Note: The marginal R-squared shows the model fit purely using the fixed-effects, and the conditional R-squared shows the model fit using mixed-effects.
• With the onset of autumn storms, the slurry-wetted pasture continued to be wetter than the semi-natural grassland, being much closer to saturation (i.e., 4.6% vs. 12.9% below saturation, respectively). These wetter antecedent conditions could mean that such pastures saturate quicker and consequentially produce more of the rapidly moving Saturation-excess Overland Flow, and so heighten downstream flood risk. Experimental research is needed to quantify if a greater mean wetness of slurry-managed pasture soils in the autumn does translate into a greater incidence, magnitude and before results and conclusions can be applied to regional-scale models.

SUPPORTING INFORMATION
Additional supporting information may be found online in the Supporting Information section at the end of this article.

MOISTURE-PROBE
A simulated semi-variogram with 2% θ V error (identical to the soil moisture-probe) was generated over a grid of equal-scale to the plots used in this study, with the error uniformly distributed. This simulated semi-variogram was used to assess the impact of the moisture-probe uncertainty on semi-variogram model parameters. The actual range in metres is given below the effective range. Note that the sill is the nugget plus the partial sill. This table should be used alongside Figure S1, and compared with Table 5 and Figure 13.