Comparing fine‐scale structural and hydrologic connectivity within unimproved and improved grassland

Grasslands vary with diverse forms and functions ranging from monocultures of perennial rye grass to more biodiverse unimproved grasslands which cover around 5% of Europe. Despite the broad diversity of grassland types, within environmental and flood risk models grasslands are frequently represented by a singular set of hydrological and structural parameters which belies their diversity and complexity. This study aimed to determine empirically the extent to which improved versus unimproved grasslands exhibit different hydrological connectivity. Working in SW England at neighbouring field sites with comparable slopes and rainfall regimes, we used unpiloted aerial vehicles to survey a tussocky Molinia caerulea dominated unimproved grassland field (MCUG) field and a Lolium perenne dominated improved grassland (LPIG) field. Using digital photogrammetry workflows applied to the overlapping aerial images, we produced a digital surface model (DSM) at 0.03‐m resolution from which flow pathways were modelled using GIS and compared with 1‐m LiDAR and DSM produced by a global navigation satellite system (GNSS). MCUG had longer, tortuous pathways through the dense tussock network with a drainage density of 2.54 m m−2. This was significantly greater than drainage density in the LPIG (1.82 m m−2). As a result of this study, we rescaled the Manning's n value for MCUG according to photogrammetrically‐derived roughness values. We suggest it should lie between 0.075 and 0.09. Our data shows that MCUG can play an important role in reducing overland flow impacts when compared to LPIG through lower connectivity which can delay run‐off to rivers.

hydrological parameter set describing grassland function. Such a simplification is problematic because there is evidence that unimproved and improved grassland exhibit significantly different eco-hydrological behaviour from each other, particularly in terms of their potential connectivity and also in terms of their run-off generation and contribution to flooding downstream . This paper will refer to 'connectivity' meaning structural connectivity, and 'disconnectivity' meaning the disruption of such connectivity in space. Connectivity herein refers to static spatial patterns in the landscape which influence water transfer and flow pathways (Bracken et al., 2013).
Improved grassland is known to be a source of substantial surface run-off due to the high levels of soil compaction, high surface and subsurface drainage connectivity and low surface roughness McIntyre & Marshall, 2010), the latter being exemplified by Manning's 'n' value for grasslands being five times lower than woodlands (Chow, 1959). Improved grasslands contribute 60% of the nitrate, 25% of the phosphorus and 75% of the suspended sediment pollution in UK rivers, costing the water industry approximately £120 million per year from the contamination of drinking water via diffuse pollution (Holden et al., 2014;Pretty et al., 2000). Conversely, some work has found that hydrological flows and connectivity can be quite different in unimproved grassland compared to improved grassland.
For example, Puttock and Brazier (2014) discovered low connectivity within mature Rhôs pasture/Culm grassland (Molinia caerulea pasture), due to the lack of field run-off synchronicity with nearby river peaks.
Corroborating this, Bond et al. (2020) found lower flow velocity in rush dominated-pastures compared to improved grassland due to denser vegetation and higher surface roughness. Despite these differences, grassland heterogeneity is rarely acknowledged in hydrological models. For example, in those which use roughness coefficients (e.g. Manning's n) to simulate surface run-off velocity and volume, areas with 'grass' land cover are distinguished only as having 'short' or 'tall' grass-which fails to acknowledge the variable hydrological processes linked to spatial changes in structure (Chow, 1959). For example Collins et al. (2012), Hansen et al. (2007) and L opez-Vicente and Alvarez (2018) all grouped grassland as a single hydrological response unit when modelling agricultural systems. There is not currently an evidence basis underpinning a more nuanced classification of 'grasslands' in such models.
As awareness of flood risk under climate change grows, different approaches to flood management are being explored at national and international levels. There is growing interest in natural flood management (NFM)-which works with natural processes to reduce flood risk by using environmentally sensitive techniques to manage sources and flow pathways of flood waters (Dixon et al., 2015). NFM also favours methods that store water in landscapes and increase hydrological residence time by reducing connectivity. However, unlike the current dominant engineering approach to flood management, NFM lacks an established evidence base describing how different techniques reduce flood risk which is needed to support future flood management decisions (Burgess-Gamble et al., 2017;Dadson et al., 2017;Ellis et al., 2021). While knowledge of the hydrological processes that reduce flood risk are well established for woodland (e.g. increased evapotranspiration) (Odoni et al., 2010;Quinn & Wilkinson, 2019;Thomas & Nisbet, 2007), grasslands have been comparatively overlooked, despite their large spatial extent. Therefore, more research into grassland eco-hydrology, spatial heterogeneity and connectivity is vital if their role in NFM is to be understood.
In Europe improved grassland is associated with intensive management including, subsurface drainage, regular fertiliser applications and high density animal grazing Laidlaw & Frame, 2013;Pilgrim et al., 2010). Unimproved grassland conversely now forms a small component of grassland coverage; covering an estimated 5% of European grassland having decreased with agricultural expansion and the need for permanent grazing pasture (Peyraud et al., 2014). Unimproved systems are highly biodiverse, owing to the lack of treatment with fertiliser and with low intensity agriculture (Blakesley & Buckley, 2016). There is a need to understand unimproved grassland hydrological structure, function and particularly structural connectivity.  Trust, 2014). These environments are known to store water above and below the surface, but very little is known of the surface hydrological processes and connectivity of these grasslands which may provide NFM benefits (Puttock & Brazier, 2014), despite a widespread goal to reinstate areas of Culm grassland for flood risk management across Devon.
Research has shown that improved grassland exhibits enhanced hydrological connectivity compared to unimproved grassland  being present towards the base of the field, the field is still classified according to the National Vegetation Classification as improved grassland by a botanical survey undertaken as part of the project.
Given the nature of the hypotheses, field sites used needed to be consistent. The MCUG and LPIG fields were chosen for their close proximity with the same underlying soil type (Wickham 2) and climatic regime. A 1-m resolution DTM derived from LiDAR was used to measure the field slopes which averaged 2-3 on both. This meant no marked differences in hydrological processes caused by slope, soil type or other climatic changes.

| Data collection
In order to collect high spatial-resolution digital surface models of the paired grasslands, an Unpiloted Aerial Vehicle (UAV) was deployed.
The DJI Mavic Air is a portable, lightweight UAV (430 g) with a high quality camera (2.3" CMOS, 12 mega pixels) which was easy to transport in the tussocky field. Three UAV flights were undertaken March,

| Data processing
The images obtained from the UAV flights over MCUG and LPIG The relative precision ratio was calculated for both field sites by dividing the UAV survey height by the RMSE (Smith et al., 2016). This measure gives a good indication of average model performance but does not identify where spatial variability is greatest or least which can be identified with precision analysis but was deemed not appropriate for this application Milan et al., 2007).
Surfer (version 13.5.583) was used for DSM generation as Agisoft Metashape does not permit the same end-user control over DSM generation compared to Surfer (Bakker & Lane, 2017;Glendell et al., 2017). Surfer allows the method of DSM generation to be chosen, and allows the user designation of the DSM pixel size. The DSM for both fields was modelled using ordinary spherical kriging (Luscombe et al., 2015).
The measurement of tussock grassland structures with remote sensing approaches has had limited exploration (Fahey et al., 1998).
Consequently, it was necessary to validate the DSM against independent measurements in order to address the proposed hypotheses.
This validation also allowed for comparison of survey techniques as per hypothesis 3. To create the validation DSM, GNSS with a spatial x, y and z accuracy of 0.02 m was used to thoroughly survey all ground and tussock points within a 3 Â 3 m area delivering 317 points in total. At every change in elevation a point was recorded on the DPGS, including multiple points across tussock tops. After post-processing using a RINEX system, these points were interpolated using ordinary spherical kriging within ESRI ArcGIS Pro (version 2.4.0). The same process of kriging was then used upon the same area of DPC generated by the UAV for comparison. Semivariogram analysis (ordinal spherical kriging) was used to characterise the spatial structure of each dataset and the semivariogram parameters compared to assess similarities between the kriged products derived from the DPC and GNSS.

| Hydrological processing
Once DSMs of each field were produced surface flow pathways could be modelled to infer function from structure. Firstly, the location and size of tussock tops were identified using a modified script from the R lidR package (version 3.0.3), a package developed for identifying tree properties within canopies from a DPC (Roussel et al., 2019) although not so far used in tussock grassland systems. The use of a DPC for tussock identification was preferable to DSM analysis as there can be error introduced around tussocks structure during the production of a DSM from the DPC (Bater & Coops, 2009). Ground points were identified and tussocks classified using a set height threshold (0.1 m above normalised ground) to identify tussocks of 0.1-0.5 m in height. From this, a raster layer describing tussock area was extracted. Trees in the LPIG and MCUG field were also segmented using the same method.
The subsequent raster layer meant that trees could be excluded from hydrological analysis. This avoided trees biasing the hydrological analysis as the software would otherwise have allowed these to impact flow pathways. Of note is the fact that trees can be chosen to be included or excluded from this workflow depending on whether they are deemed to exert important controls on flow paths. The standard deviation (σ) of z points of the DPC was plotted in Surfer (version 13.5.583) across the whole of each field as a proxy for surface roughness as curved and rougher surfaces will have greater σ in the z coordinate .
To perform surface water hydrological analysis, the DSM needed to be filled for sinks or data gaps to predict potential flow pathways through the specified area. ArcGIS fill function has been shown by Venticinque et al. (2016) and Pareta and Pareta (2012) to be suitable for filling DSM gaps in large river basins. However, in a landscape such as a tussock field we felt that adopting this method could have resulted in loss of valuable depressions which we hypothesised were essential structural attributes that impact surface water storage in MCUG (Jenkins & McCauley, 2006). Instead, the optimised pit removal function uses a combination of cut and fill to minimise loss of fine topographic features within this landscape (Soille, 2004).
This function was applied to both the experimental field and the control field.
The hydrological toolset within ESRI ArcGIS Pro (version 2.4.0) was used for flow pathway replication. Firstly, flow direction of every cell within the raster was computed from the DSM in D8 directions.
Flow accumulation then identified areas of accumulation into each downslope raster cell, which was then reclassified to reflect accumulation of different drainage areas (10,20,50,100,200,500 and 1000 m 2 ). The stream order function produced a raster of flow pathways based on drainage areas, and the same process was applied to a 1-m LiDAR DSM for comparison with commonly used data in hydrology (Barber & Shortridge, 2013). Drainage density (Dd) was calculated using the line statistics function and was calculated as the total length of flow pathway per unit area. This parameter can be used as a proxy for connectivity within a field (Godsey & Kirchner, 2014).
Each field was divided into a grid (1, 4, 25, 100, 400, 1225 and 2500 m 2 ) within which flow pathway length was calculated. This meant that fields of different sizes could be compared for connectivity parameters; higher Dd values indicated lower connectivity as more pathways were present. Each grid was then compared using a Welch two-sample T-Test to test for significant difference in flow pathway length and therefore connectivity. A Welch two-sample T test was used as the data were normally distributed with independent samples.
Flow pathway values were then extrapolated using flow pathway averages (±1 standard deviation) to provide a rough prediction of the flow pathway length in a larger M. caerulea and LPIG field up to 50,000 m 2 . A summary of the entire data processing workflow is summarised in Figure 2. 3 | RESULTS

| DSM
The DSM of the MCUG field and LPIG field are shown in Figure 3. F I G U R E 2 Study workflow

| Validation
The results of the GNSS and SfM photogrammetry comparative 3 m Â 3 m plots are shown in Figure 5. Both show the basic tussock structure of the landscape. The points were interpolated using ordinary spherical kriging, the semivariogram generated for the kriged points are shown in Figure 6 and Table 2 Table 2). The GNSS DSM was subtracted from the SfM photogrammetry DSM to understand differences in the two methods. The two DSMs show that tussocks are still fundamentally located at the same locations, though some variation in height and location exist and contributes to differences in semivariance parameters (Table 2).
There is an average x/y/z error of À0.10 m between the GNSS and DPC. This may reflect small x/y error within the UAV DSM, suggesting   Table 2 shows that the UAV is a validated close representation of tussock reality and therefore valid conclusions of flow pathways can be made from the SfM product.  increased with area in MCUG as shown in Figure 8. The use of a Welch two sample T-test showed there was no difference in mean flow path length between the two fields below 1225 m 2 (p = 0.041, ≥0.95) (Table S1). There was also a significant difference at 2500 m 2 (p = 0.004, ≥0.95). When extrapolated using flow pathway averages up to 50,000 m 2 (5 ha), the difference in flow pathway length continues to increase, as shown in Figure 8. Upper and lower limits of Dd were produced by calculating minimum and maximum flow pathways per unit area from observed flow pathway data. The contrast in Dd between unimproved and improved grassland cover increases with area, and the gap between the two vegetation types continues to increase. Figure 9 shows the SfM derived drainage density plotted against drainage density of the LiDAR DSM flow pathways, from which a notable difference between the two methods can be seen.  Cammeraat and Imeson (1999) and Quinton and Carey (2008). The study advances understanding of surface processes beyond the plot-scale soil moisture plots of unimproved grassland work by Ludwig et al. (2005) and Wallace and Chappell (2020) by including surface flow pathways over a field extent. We also demonstrated that the significant difference in flow pathway length between unimproved and improved grassland increased with area (Figure 8),

| Flow pathways
which we tested to a maximum extent of 5 ha. This difference suggests that over larger areas of unimproved grassland, potential flow attenuation due to flow pathway length increasing will grow.
How do tussocks impact surface flow pathways?
Our work has proven that SfM photogrammetry offers a capable and scale-appropriate method for describing surface structures that impact hydrological surface flow pathways. Z point error indicated that MCUG had a significantly rougher surface than LPIG (0.03 m versus 0.01 m) (Figure 4). Surface roughness influences connectivity in fields by increasing flow pathway length and creating surface depressions for water storage (Bracken & Croke, 2007). In response to hypothesis 1, we showed that the MCUG field was significantly rougher than the LPIG field through z point error (σ), as rougher surfaces produce more z error on nadir UAV flights than planar surfaces Shepard et al., 2001). Point error was on average 200% greater in MCUG fields than LPIG fields. As a result of this finding, we recommend the Manning's n for rough, tussock structure grasslands be set to reflect this 200% roughness increase to a value range of between 0.075 and 0.09. This would be suitable in comparison to the less rough improved grassland currently classified at 0.025 to 0.03.
As suggested by Dadson et al. (2017), the lower connectivity in complex environments has potential NFM benefits when soils become saturated by prolonged rainfall or during high intensity events with infiltration excess overland flow. When overland flow is generated, the same volume of water has further to travel through the tussock network in the unimproved grassland, than in the better-connected improved grassland. Although hydrographs were not measured here, other work from woodland ecosystems shows that where flow pathways through landscapes become more disconnected and friction due to surface roughness increases, the peak of the hydrograph is delayed (Papanicolaou et al., 2018;Thomas & Nisbet, 2007). There are also further benefits-for example, MCUG fields will have less soil erosion due to the more sinuous pathways which is otherwise a common problem and cause of land degradation in poorly managed improved grassland fields (Brazier et al., 2007). These pathways have been observed in other tussock grassland; in one study of wetland sedges and Carex stricta tussocks, tussock structure microtopography made an ideal habitat for a variety of other plant species to colonise which increased surface roughness and sediment trapping (Werner & Zedler, 2002). We argue with the evidence gathered in this work, that M. caerulea-dominated unimproved grassland has great potential as a form of NFM through delivering increased surface roughness and more sinuous flow pathways compared to improved grassland. There are also multiple further environmental benefits from unimproved grassland restoration such as increased flora and fauna biodiversity and carbon sequestration (Puttock & Brazier, 2014).
Do unimproved and improved grasslands have significantly different drainage density?
We found a significant difference in drainage density at 1225 m 2 between MCUG and LPIG, the difference increasing with area ( Figure 8). Evidence from Figure 9 and T test of mean flow pathway length per unit area shows that the NFM benefit of MCUG was marginal over smaller spatial extents than 1225 m 2 . Therefore, larger areas of MCUG would likely be needed to deliver significant hydrological benefits via increased flowpath lengths over LPIG. Culm grassland, which includes tussocks of M. caerulea, is currently being restored across fragmented areas of grassland in the South West of the UK, with a stated goal of increasing water storage and attenuation in restored fields (Devon Wildlife Trust, 2014). At present, areas of unimproved grassland are often over very limited extents (<2000 m 2 , i.e., small fields) and fragmented (Blakesley & Buckley, 2016;Devon Wildlife Trust, 2014). This research highlights that more extensive areas should be restored if NFM goals are to be achieved, particularly as at present, fields of Culm grassland are highly fragmented with no landscape connectivity and thus less chance to impose more hydrological disconnectivity. The fragmentation of wetlands reducing their potential hydrological benefits has been highlighted globally, such as China and the USA (Johnson et al., 2014;Liu et al., 2020). Results here suggest that restoring wetlands or unimproved grasslands for hydrological reasons (e.g., reducing hydrological connectivity) needs to be done on spatial extents exceeding at least 1225 m 2 .
The hydrological impact of microtopographic features such as tussocks acting over large areas >1225 m 2 can deliver important functional shifts within catchment systems (Antoine et al., 2009;Phillips, 1988). Phillips (1988) states that geomorphic systems are characterised by complex process interactions at multiple scales, of which microtopography can be one such influence upon hydrological processes such as infiltration and overland flow pathways, but often with limited understanding. Dunne et al. (1991) argues that infiltration rates on grassland slopes vary with the interaction of rainfall, run-off and vegetation topography because microtopography increases the surface area over which infiltration can occur. The importance of microtopography in controlling connectivity that results in NFM benefits is not widely quantified, particularly in grasslands, even though some analogous environments have been studied. For example, Courtwright and Findlay (2011) found microtopographic indentations in the tidal swamp of the Hudson River created areas of water storage which altered the basin's storage capacity. The duplication of the mounds and pooling in forested wetlands was used by Barry et al. (1996) to restore wetland properties of water storage according to a similar mechanism observed here in M. caerulea-dominated unimproved grassland. Appels et al. (2011)  Three different survey methods were deployed within this research: SfM photogrammetry, LiDAR and field assessment using a GNSS. SfM photogrammetry provided the finest spatial resolution product with which to assess differences in microtopography and subsequently to quantify potential impact upon connectivity between unimproved and improved grassland. LiDAR is a common data source for hydrological modelling, with resolution up to 0.25 m possible in the UK to deliver high resolution catchment assessments (England & Gurnell, 2016;Vierling et al., 2008). The LiDAR DSM used for analysis in this study had a relatively high resolution at 1 m, a size which has been used to assess wetland connectivity, such as Wu and Lane's (2017) assessment of wetland depression ability to disconnect flow pathways. Nonetheless, we found that this spatial grain is inadequate to capture the fine microtopography present in unimproved tussock grasslands which clearly influence field connectivity (Figures 7 and 9 The use of a GNSS to survey points manually to form a DSM could be considered an alternative method to LiDAR, particularly when greater accuracy in three dimensions is needed. For example, Higgitt and Warburton (1999) and Nuimura et al. (2012) both used GNSS to assess meander changes in remote upland areas that lacked fine-spatial resolution LiDAR. The GNSS in this study provided accurate XY points from which to build a DSM through kriging ( Figure 5), but a point density of 38.44 points per m 2 was insufficient to capture the tussock outline in comparison to SfM photogrammetry. The mean z error of À0.11 m may reflect average vegetation height upon the tussocks as they regrew, as the GNSS rover recorded tussock height from soil mass, not from small amount of grass height upon the tussock which had grown after swaling. In contrast, a UAV DPC will incorporate grass upon the tussock as tussock structure. This error is minor when the tussock structure was still clearly captured. The UAV survey had 1493 points per m 2 which was more than adequate to capture the structures in detail for hydrological analysis. SfM photogrammetry to assess vegetation structures has already been proven as a strong method in ecology, such as for vegetation surveys and biomass quantification (Cunliffe et al., 2016;Dandois & Ellis, 2010;Puliti et al., 2015) and is being explored as a method to assess fluvial features (Debell et al., 2016;Woodget et al., 2015).

| CONCLUSION
Unimproved grasslands with microtopographic tussock structures of M. caerulea can play an important role in reducing overland flow impacts when compared to improved grasslands. The lower connectivity of these environments may reduce overland flow by storing and delaying run-off to rivers. The potential benefits of disconnectivity have been shown to increase with area, making a case for restoring unimproved grasslands on a large scale. The benefits of unimproved grassland for flow attenuation can be used widely through the restoration of sites that are currently improved grassland.
The use of SfM photogrammetry to assess connectivity was highly effective for grasslands when compared to traditional methods of both LiDAR and manual surveys using a GNSS. The use of UAVs in NFM should therefore be considered as a method to assess field-scale structure and function for vegetation with limited understanding, such as those within unimproved grasslands. This understanding can then be used to understand the impact of field connectivity across a catchment, adapting roughness coefficients and field properties accordingly within models. This study also demonstrates the importance of not treating grasslands as one homogenous unit in hydrological modelling, and parameters such as roughness coefficients should be adjusted by researchers accordingly. The potential of SfM photogrammetry to assess unimproved grassland with complex microtopography such as M. caerulea tussocks is substantial.

ACKNOWLEDGEMENTS
We thank our reviewer Professor Andy Baird whose comments helped improve and clarify this manuscript. We would also like to thank the multiple colleagues who supported fieldwork. We acknowledge funding support from the Environment Agency, Devon Wildlife Trust, INTERREG project (Climate resilient community-based catchment planning and management) Triple-C and the University of Exeter.

CONFLICT OF INTEREST
The authors have no conflict of interest to declare.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available at DOI, with code available online (at https://github.com/exeter-creww/Ellis_ et_al_2021_supporting_information).