j ourna l homepage: www.e lsev ie r .com/ locate /geomorph
Floodplain complexity and surface metrics: Influences of scaleand geomorphology
Murray W. Scown a,⁎, Martin C. Thoms a, Nathan R. De Jager b
a Riverine Landscapes Research Laboratory, University of New England, Armidale, NSW 2350, Australiab United States Geological Survey, Upper Midwest Environmental Sciences Center, La Crosse, WI 54603, USA
⁎ Corresponding author at: 26 W. Martin Luther Kin45220, USA.
Many studies of fluvial geomorphology and landscape ecology examine a single river or landscape, thus lackgenerality, making it difficult to develop a general understanding of the linkages between landscape patternsand larger-scale driving variables. We examined the spatial complexity of eight floodplain surfaces in widelydifferent geographic settings and determined how patterns measured at different scales relate to different envi-ronmental drivers. Floodplain surface complexity is defined as having highly variable surface conditions that arealso highly organised in space. These two components of floodplain surface complexity were measured acrossmultiple sampling scales from LiDAR-derived DEMs. The surface character and variability of each floodplainwere measured using four surface metrics; namely, standard deviation, skewness, coefficient of variation, andstandard deviation of curvature from a series of moving window analyses ranging from 50 to 1000 m in radius.The spatial organisation of each floodplain surface was measured using spatial correlograms of the four surfacemetrics. Surface character, variability, and spatial organisation differed among the eight floodplains; and random,fragmented, highly patchy, and simple gradient spatial patterns were exhibited, depending upon the metric andwindow size. Differences in surface character and variability among the floodplains became statistically strongerwith increasing sampling scale (window size), as did their associations with environmental variables. Sedimentyieldwas consistently associatedwith differences in surface character and variability, as wereflowdischarge andvariability at smaller sampling scales. Floodplain width was associated with differences in the spatial organisa-tion of surface conditions at smaller sampling scales, while valley slope was weakly associated with differencesin spatial organisation at larger scales. A comparison of floodplain landscape patterns measured at differentscaleswould improve our understanding of the role that different environmental variables play at different scalesand in different geomorphic settings.
Floodplain surfaces are spatially complex geomorphic landscapesthat display a high diversity of surface features and landforms (Thoms,2003). This spatial complexity reflects an array of environmental driversoperating at different scales and influences the hydrodynamics, ecologi-cal productivity, and resilience of these landscapes (Salo et al., 1986;Hughes, 1997). River scientists and landscape ecologists have limitedgeneral theories about how floodplain landscapes are structured, howthemeasurement of structure depends on scale, and how different phys-ical and biological variables influence floodplain landscape structure atdifferent scales. One of the primary reasons for this is limited comparabledata among different floodplain landscapes. As a consequence, manystudies focus on a single floodplain, thus lack generality. Conclusionsdrawn from such studies can only apply to the landscape studied. As
g Dr., MS 587, Cincinnati, OH
n).
comparable landscape-scale data become increasingly available, con-trasting patternsmeasured at different scales in different geomorphic set-tings is possible, providing an opportunity to investigate more generaltheories about the structure and function of these ecosystems.
Floodplains may contain a range of structural conditions or features,which may cluster into distinct physical landforms or patches, or bearranged continuously along gradients. These two properties of com-plexity are important, as a larger number of features within a system in-crease the number of potential states and interactions, while theirorganisation influences the flow of energy and material. Consequently,these two main aspects of spatial complexity (i.e., the number anddiversity of features and their spatial organisation) determine not onlyhow energy and materials are processed by systems but also how non-linear responses to inputs may emerge through feedbacks within land-scapes like floodplains (Simon, 1962; Heylighen, 1989; Phillips, 2007).Previous measurements of the spatial complexity of floodplains havelargely focused on the former of these aspects, with limited appreciationof the latter (Scown et al., in press). Advances in data-capture technolo-gies, such as light detection and ranging (LiDAR), now enable the
103M.W. Scown et al. / Geomorphology 245 (2015) 102–116
character and variability as well as the spatial organisation of floodplainsurface conditions or features to bemore easily quantified. Patternswith-in such data can be quantified using surface metrics (McGarigal et al.,2009), which indicate numerous aspects of the character and variabilityof surfaces, while the spatial organisation of conditions or features canbe measured using indices of spatial autocorrelation and geostatisticalmodels (Rossi et al., 1992; Radeloff et al., 2000; Legleiter, 2014). Thesemodels indicate the strength and characteristic length scale of organisa-tion within a landscape. Strong spatial organisation exists when nearbysamples are more highly correlated than more distant samples, whilethe length scale over which this organisation occurs is reflected by thedistance at which metrics switch from being correlated to uncorrelated(Sokal, 1979; Radeloff et al., 2000).
Scale is fundamental to observing geomorphic systems (Phillips,1988). Structuring processes exert their influence over defined spatialand temporal scale ranges, producing landscape patterns that emerge atrelative spatial scales and respond over relative time periods (Schummand Lichty, 1965). Consequently, scales of observation, or limited aware-ness of scale, can lead todifferent conclusions about pattern–process link-ages in geomorphology (Church and Mark, 1980; Phillips, 1988). Spatialscale issues are inherent in any landscape study, and these relate to thegrain (resolution of observation) and the extent of the landscape beingstudied (total area of observation) (Kotliar and Wiens, 1990; Wu andQi, 2000; Dollar et al., 2007). Changes in grain and extent are known toaffect landscape metrics (Wu, 2004). High resolution, quantitative dataprovided by LiDAR are now available over entire floodplain landscapeextents and at increasingly fine grains. Little reason to alter the grain orextent of such data exists, as this would only lead to a loss of quantitativeinformation. However, changing the sampling scale, which is akin tochanging the size of the window within which landscape information isinterpreted or filtered, will have an effect on the results observed (BarMassada and Radeloff, 2010; Ren et al., 2013). This has particular conse-quences for measuring the spatial complexity of floodplain surfacesgiven that most research has traditionally been focussed at smallsampling scales (Scown et al., in press).
Various perspectives of how floodplain surface morphology isprimarily determined exist. However, these are all centred aroundeither the availability of sediment and its character, the nature of the val-ley in which this sediment accumulates, and the energy conditions of themain river system transporting sediment. Nanson and Croke's (1992)floodplain classification of three geomorphological classes, with thirteenorders and suborders, is primarily based on stream power and sedimentcharacter. Others have suggested that boundary conditions set by the val-ley, withinwhich the flow regime and sediment delivery operate, is also aprimary determinant of floodplain morphology and processes (Hynes,1975; Thoms et al., 2000; Ibisate et al., 2011). Larger spatial- andtemporal-scale factors such as geomorphic history and physiography ofthe landscape, have also been viewed as the primary determinants offloodplain surface morphology (Warner, 1992; Church, 2002; Hudsonet al., 2008; Notebaert and Piégay, 2013). All of these levels of influenceare related; however, we suggest that confusion about which variablesare most important in structuring floodplain environments is owing tothe fact that the spatial patterns observed depend on the scale at whichthey are measured. Further, associations between floodplain structureand external variables have, to date, largely been qualitative. Wehypothesise that the driving variables most important for determiningthe two components of spatial complexity in floodplains will dependupon the spatial scale of measurement and that particular variables willnot be the main drivers of both components of spatial complexity norover all measurement scales. If this is the case, then it should be possibleto develop a more synthetic understanding of the role different variablesplay at different scales and in different geomorphic settings.
The aim of this study is twofold. First, the two components of spatialcomplexity are quantified using moving windows of different sizes foreight floodplain surfaces from different geographic regions; and second,statistical associations between scale-specific landscape patterns and a
set of environmental drivers previously hypothesized to influence flood-plain surfaces are established. Three questions are raised: (i) Does surfacecharacter and variability differ among floodplains, and which surface at-tributes are these differences associatedwith? (ii) Does the spatial organi-sation of surface conditions differ among the eight floodplains, andwhichspatial model parameters are these differences associated with? (iii)Which environmental variables are associatedwith the differences in sur-face character and variability, and with differences in the spatialorganisation of surface conditions? These questions are investigatedover three sampling scales.
2. Study area
Eight floodplains from different geographic settings were selected forstudy (Fig. 1) based on the availability of high-resolution, LiDAR-deriveddigital elevation models (DEMs). Six floodplains were located inAustralia (Gwydir, Macquarie, Murrumbidgee, Narran, Woodforde,Yanga), one in the United States of America (Upper Mississippi), andone in South Africa (Shingwedzi). Details of the eight floodplains are pro-vided in Table 1; and in summary, they differ in terms of their degree ofvalley confinement, climate, and position within the stream network(Table 1). Four of the floodplains studied are contained within relativelyconfined river valley troughs. These are (i) the floodplain of theShingwedzi River, located at the confluence of the Shingwedzi andMpholonga Rivers in Kruger National Park, Limpopo, South Africa(31°24′E, 23°05′S); (ii) the floodplain of the Woodforde River, locatedin central Australia, ~ 150 km north of Alice Springs, Northern Territory(133°20′E, 22°21′S); (iii) the floodplain of the lower MurrumbidgeeRiver (‘Bidgee’), between the town of Balranald and its confluence withthe River Murray in New South Wales, Australia (143°24′E, 34°42′S);and (iv) the floodplain of the UpperMississippi River, locatedwithin nav-igation pool 9 (91°15′W, 43°29′N), between the states ofWisconsin,Min-nesota, and Iowa, USA.
In contrast, the remaining four Australian floodplains are allcontained within relatively unconfined valleys that are up to 60 km inwidth, within the Murray Darling basin. The Narran floodplain (v) is aterminal floodplain located at the end of the Narran River, ~ 75 kmwest of Walgett in New South Wales, Australia (147°23′E, 29°48′S).The Narran River is one of four major anabranch channels, whichmake up the lower Balonne River system. The Yanga floodplain (vi)located upstream of the Bidgee floodplain in New South Wales,Australia (143°42′E, 34°30′S), is an unconfined valleywhere floodwaterscan span up to 20 km. The floodplains of the Macquarie River (vii;147°33′E, 30°41′S) and lower Gwydir River (viii; 149°20′E, 29°16′S) arelocated in geomorphological zones containing numerous anabranchchannels and floodplains up to 60 km in width that are heavily dissectedby distributary channels.
3. Methods
3.1. Digital elevation models
Bare-earth digital elevation models (DEMs) derived from airbornelaser scanning (LiDAR) were used to represent the topographic surfaceof each of the eight floodplains. These were obtained from various gov-ernment and nongovernment agencies and have undergone rigorousquality assurance protocol. From each original DEM, the floodplainextent was manually delineated to remove any valley side effects. Thisdelineation was based on examination of breaks of slope in the DEM,contours, changes in vegetation from aerial photography, and floodwaterextents derived from Landsat TM imagery. A buffer within this manuallydelineated extent was also removed to ensure that nothing other thanwhatwas deemed to be part of the floodplainwas included. Permanentlyinundated areas were removed because of the difficulties in attaining ac-curate subsurface land elevations using LiDAR. Areas of development,farm dams, artificial levees, and irrigation channels were also removed
Fig. 1. Digital elevation models displaying the floodplain surface in metres above sea level for each study site (A) to (H) as listed in Table 1. Crosses indicate coordinates listed in text.
104 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
from the DEMs so that as close to the ‘natural’ floodplain surface aspossible was obtained. Each DEM was then detrended by generalisingthe 1-m valley contours across the floodplain and creating a trendsurface—based on the downstream slope—in order to remove the overalldownstream slope. This trend surfacewas then subtracted from the orig-inal DEM to produce a detrended DEM containing heights above a zero
Table 1Summary of the geographical settings and stream gauging stations of the eight study floodplai
Floodplain name Valley setting Climate Network setting
Highly confined with a valley trough generally between 1 and 3 km in width. Overalloften N10 m below the floodplain surface. Relatively shallow flood channels with a b
(B) ‘Woodforde’ Confined Arid Headwaters continConfined within a valley trough up to 3 km in width. The floodplain surface is incised bpunctuated by crevasse splays in places.
(C) ‘Bidgee’ Confined Semiarid/temperate Lowland continuouValley trough generally around 4 km in width, confined between aeolian ridges andfloodplain contains a highly sinuous channel, with meander scrolls and numerous bi
(D) ‘Mississippi’ Confined Continental Upland continuousConfined between dolostone-capped sandstone bluffs with a valley trough up to 6 kmUpper Mississippi and dominant geomorphic features include islands, natural levees, c
(E) ‘Narran’ Unconfined Semiarid Lowland terminalTerminal floodplain system dominated by a large central lake with two smaller lakeplanform before entering these lakes.
(F) ‘Yanga’ Unconfined Semiarid/temperate Lowland continuouHighly sinuous main channel, with shallow reticulate channels dissecting much of theaeolian dunes form high points throughout the floodplain.
(G) ‘Macquarie’ Unconfined Semiarid/temperate Lowland continuoMany shallow channels dissect the surface and the floodplain is dominated by a reti
(H) ‘Gwydir’ Unconfined Semiarid/temperate Lowland terminalRelatively flat and featureless over large areas; several shallow channels dissect the
datum, which represented the lowest point relative to the overall down-stream slope in the floodplain. For Shingwedzi, the 2.5-m contours wereused because of its steeper nature;while forMississippi, the trend surfacewas based on the mean low water surface elevation. Finally, thedetrended DEMs were all resampled to a 5 × 5 m grid size using thecubic method in ArcGIS 10.2 to ensure consistent resolution among
ns; a brief geomorphological description is also included below each floodplain.
Gauging station Agency
Shingwedzi River @ Kruger National Park (B9H002)Mpholonga River @ Kruger National Park (B9H004)
SA DWAF
riverine landscape characterised by a large, deeply incised main channel whose bed israided pattern dissect the floodplain surface.uous Murrumbidgee River @ D/S Balranald Weir (410130) NSW DPIy a series of small, braided flood channels and a well-developed levee system, which is
s Mississippi River @ McGregor, IA (05389500) USGStertiary-aged terraces that rise between 5 and 10 m above the active floodplain. Thellabongs as the dominant physical features.
Woodforde River @ Arden Soak Bore (G0280010) NT DLRMin width. Numerous anastomosing channels dissect the floodplain in this section of therevasse splays, backwaters and swamps.
Gwydir River @ Millewa (418066) NSW DPIs on the distal parts of the floodplain. The main channel dissipates into a reticulate
s Macquarie River @ Gibson Way (421116) NSW DPIfloodplain surface. Numerous small terminal lakes occur at distal regions. Occasional
us Narran River @ Wilby Wilby (422016) NSW DPIculate planform.
Murrumbidgee River @ D/S Balranald Weir (410130) NSW DPIfloodplain flowing to a number of small terminal wetlands.
105M.W. Scown et al. / Geomorphology 245 (2015) 102–116
data sets and to reduce fine-grained ‘noise’ that can be evident in LiDARdata.
3.2. Surface character and variability
To quantify the character and variability of the eight floodplainsurfaces, four surface metrics were calculated for each of the detrendedDEMs using moving windows at three window sizes. These were SD,Skewness, CV, and SDCURV calculated within circular windows of radiusequal to 50, 200, and 1000m. Each of the four surfacemetrics measuresa different attribute of the floodplain surface. The standard deviation ofsurface heights (SD) indicates surface variability about the mean eleva-tionwithin an area. Skewnessmeasures the skewness of the distributionof surface heights, which indicates peak and valley characteristics.Positive skewness may indicate that the surface has relatively high, in-frequent peaks or that channels and depressions are filled in. Negativeskewness may indicate that peaks are flattened or that channels areheavily incised into the floodplain. Geomorphologically, these mayreflect areas of deposition and erosion, respectively. The coefficient ofvariation of surface heights (CV) indicates surface variability relativeto the mean elevation within an area. It is useful for low-lying areaswhere standard deviation is relatively low but small elevation changesare ecologically important. The standard deviation of total surface cur-vature (SDCURV) indicates how variable curvature is across the surfaceand subsequently how convoluted that surface is. The moving windowanalyses measure each metric at each cell in the DEM based on thevalues of all cells within a radius around the focal cell. The output is asurface model containing the specific surface metric values measuredusing a specific window size centred on each cell.
3.3. Spatial organisation of surface character and variability
To quantify the spatial organisation of surface character and variabil-ity within each floodplain at each scale, 1000 random sample sites wereselected within each floodplain. These sites had no requirements interms of their minimum or maximum distance apart in order to gain agood sample of pairs of sites over all spatial lag distances. At each of thesample sites, the four surface metric values at three window sizes wereextracted and range standardised. Moran's I was then calculated locallyfrom the range standardised samples using the ncf package in R. Moran'sI is an index of spatial autocorrelation in a variable in which a value of 1indicates perfect spatial correlation, while a value of−1 indicates perfectspatial dispersion. A value of zero indicates a random spatial distributionin the variable. Local calculation of Moran's I enables changes in spatialautocorrelation with increasing distance to be determined. A range oflag distances from 100 m up to half of the maximum distance betweensamples within a floodplain were used at equal intervals of 100 m. Anylag distanceswith b 200 pairs of sampleswere removed to avoid spuriousMoran's I values for comparatively low number of pairs (most lag dis-tances had more than 1000 pairs).
Each Moran's I was then plotted against lag distance to create aspatial correlogram. The spatial correlogram shows changes in spatialautocorrelation with increasing distance between samples (Sokal,1979; Radeloff et al., 2000). Exponential isotropic models were fittedto the correlograms using nonlinear least squares regression in thestats package in R. The exponential isotropic model is defined as:
I ¼ C0 þ C 1−e−hA0
� �ð1Þ
where I is Moran's I for lag distance h, C0 is the nugget autocorrelation ath = 0, (C0 + C) is the maximum spatial dispersion in the sample ofsurfacemetric values, and A0 is the range parameter so that the effectiverange (3A0) is the distance at which (C0 + C) is within 5% of the asymp-tote (Robertson, 2008). The nugget (C0) indicates the strength of orga-nisation or patchiness of the surface metric, with values approaching
(or exceeding) 1 indicating a highly organised spatial distribution andvalues approaching 0 indicating amore randomarrangement. AlthoughmeasuredMoran's I ranges between−1 and 1, modelled nugget valuesmay exceed 1 if the model rises steeply with lag distances approachingzero. The range (A0) indicates the distance atwhich spatially distributedsurface metric values become uncorrelated; that is, the characteristiclength scale of organisation or patches. The nugget and range parametersof the best fitting exponential isotropic model were automaticallyextracted once the model had converged on a solution. Standard errorsof the model parameter estimates were also extracted from the modelsolution.
3.4. Multivariate statistics
To quantify differences in surface character and variability betweenthe eight floodplains, 50 random sample sites were selected withineach floodplain. These sites were set at a minimum distance of 500 mapart to reduce overlapping of the large window sizes, and allcontained a minimum of 60% data in the original DEM. Obtaining50 random samples with large windows containing 100% data wasnot possible in most of the floodplains, therefore this trade-off wasdeemed appropriate. At each of the sample sites, the four surfacemetricvalues at three window sizes were extracted.
The floodplain samples were then ordinated in multidimensionalspace for each window size using the Gower dissimilarity coefficientthat is range-standardised to account for differences in surface metricranges. Permutational multivariate analysis of variance (PERMANOVA)(Anderson, 2001)was then conducted in Primer 6 to determinewhethersignificant differences in surface character and variability betweenfloodplains—based on the four surface metrics within each windowsize—occurred. The degree of pairwise separation between floodplainswas then determined using analysis of similarity (ANOSIM) (Clarke,1993) in Primer 6. ANOSIM provides a value between 0 and 1, with0.00–0.25 considered no separation, 0.26–0.50 weak separation,0.51–0.75 moderate separation, and N 0.75 strong separation, as a ruleof thumb. Associations between surface metrics and multivariate pat-terns of separation between floodplains were then determined usingthe envfit routine in the R package vegan. Envfit provides a vector foreach variable that maximises correlation between that variable andthe projections of ordination points onto that vector. Significance isassessed using permutation of variables.
Groups of floodplains with similar spatial organisation of surfacecharacter and variabilitywere thendetermined using hierarchical cluster-ing in R. The nugget and range parameters for each of the surface metricswere used as the eight spatial organisation variables for each floodplainwithin each window size. Again, the Gower dissimilarity coefficient wasused, followed by clustering using the flexible unweighted pair groupmethodwith arithmeticmean (UPGMA). The optimumnumber of groupswas determined by the first major inflexion point on the scree plot ofnumber of groups versus level of association, as advised by McGarigalet al. (2000). Associations between spatial organisation variables andmultivariate patterns of separation between floodplainswere then deter-mined using the envfit routine in R.
3.5. Associations between environmental variables and components ofsurface complexity
To investigate associations between potential environmental driversand the two components of surface complexity, seven environmentalvariables were calculated for each of the floodplains. These variablesreflected discharge volume and variability, energy conditions, degreeof valley confinement, and sediment availability. The four discharge var-iables were mean daily discharge in ML/day (Q), CV daily discharge(QCV),CVmean annual discharge (QCVAnn), and CVmaximumannual dis-charge (QCVMax). The energy and confinement variables were average
106 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
valley slope in m/m (Vs) and average floodplain width in km (Fpw).Sediment availability was calculated as sediment yield in t/km2/y (SY).
The discharge variables for each of the floodplains were calculatedfrom daily flow data for the period 1 December 1983 to 31 January2013. This was the longest period for which continuous reliable flowdata were available for each of the rivers. The gauging station nearestto each floodplain, which provided reliable data over this period, wasused (Table 1). For the Shingwedzi, the average of the hydrologicalvariables from each of the two river gauging stations was used.
The energy variables were calculated manually in ArcGIS 10.2 foreach of the floodplains. Average valley slope was measured along thevalley centre line, which was manually digitised upon the originalDEM for each floodplain. Floodplain width was approximately mea-sured at five equidistant intervals along the valley centre line based ona visual examination of the DEM, floodwater extent derived fromLandsat TM imagery, and vegetation changes from aerial photographyand satellite imagery; the average of these five measurements wastaken as the average floodplain width.
Sediment yield was estimated from the best available sources foreach of the floodplains. For the Shingwedzi, sediment yield was esti-mated using the midpoint of the regional range reported by Msadalaet al. (2010) in their South African sediment yield predictions. For theBidgee, Gwydir, Macquarie, Narran, and Yanga, the midpoint of theriver sediment load ranges reported by De Rose et al. (2003)was dividedby catchment area to provide a yield estimate. For the Mississippi, theaverage of the sediment yield values for the Upper Mississippi Riverreviewed and reported by Turner and Rabalais (2004) was taken.For the Woodforde, the midpoint of the sediment yield estimaterange for this region reported by Lvovich et al. (1991) was used. Allenvironmental variables were normalised using a base 10 logarithm
Fig. 2. Distribution of surface metric values for each of the floodplains at each measurement
and then standardised against their maximum before subsequentanalyses.
Associations between the seven environmental variables and thetwo aspects of surface complexity were then determined using theGower dissimilarity coefficient and the bioenv routine in the R packagevegan. Bioenv determines the correlation between dissimilarities offloodplain surface variables and distances between floodplains basedon environmental variables (Clarke and Ainsworth, 1993). A scoreapproaching zero indicates poor association between the optimumcombination of environmental variables and the patterns of floodplainsobserved, while a score approaching 1 indicates a strong association.The bioenv routinewas conducted for all seven environmental variablesand both the surface metric values and the spatial organisation vari-ables. Associations between environmental variables and multivariatepatterns based on both surface metrics and spatial model parameterswere then determined using the envfit routine in R.
4. Results
4.1. Surface character and variability
Surfacemetric valueswere highly variable, reflecting thewide rangeof physical character both within and between the eight floodplains(Fig. 2). The Shingwedzi had the highest median SD at all window sizesand the highest median SDCURV at the 200- and 1000-m window sizes(Fig. 2A and D). These two metrics for the Shingwedzi were also amongthemost variable of all surface metrics and themain determinant of sep-aration among floodplains in multivariate space (Fig. 3). High SD valuesindicate high surface variability, which, in the Shingwedzi was associatedwith the presence of incised floodplain channels. The Bidgee floodplain
scale from the 50 random sample sites for (A) SD, (B) CV, (C) Skewness, and (D) SDCURV.
Fig. 3. Example of the spatial organisation of surface metric values inWoodforde for (A) Skewness at the 50-mwindow size; (B) CV at the 200-mwindow size; and (C) SD at the 1000-mwindow size. Spatial correlograms and models are shown below with Moran's I from −1 to 1 on the y-axis and lag distance on the x-axis.
107M.W. Scown et al. / Geomorphology 245 (2015) 102–116
had the second highestmedian SD, at the 200- and 1000-mwindow sizesand was associated mainly with the presence of meander scrolls.Woodforde and Mississippi had the third and fourth highest median SDat the 200- and 1000-m window sizes. Values of SD for the Narran,Yanga, Macquarie, and Gwydir were relatively low by comparisonat all window sizes (Fig. 2A), reflecting a lower topographic relief.Overall, the range of SD and SDCURV values indicate the four confinedfloodplains had a more variable surface elevation than the four uncon-fined floodplains.
The Mississippi, Macquarie, and Yanga floodplains had three of thefour highest median SDCURV values at all window sizes; while theNarran, Gwydir, and Woodforde always had three of the four lowest(Fig. 2D). High SDCURV values indicate a highly convoluted surface. TheGwydir and Narran had among the lowest median SD and SDCURV of allfloodplains at all window sizes. They also had a median Skewness ofapproximately zero at all window sizes. This reflects that these twofloodplains are extremely flat over large areas relative to most of theother floodplains. The Gwydir also had low CV over all window sizes.
108 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
Patterns of CV and SDwere similar in theWoodforde, Bidgee, Yanga,and Gwydir floodplains indicating that local SD is likely related to localmean surface height in these floodplains. The Shingwedzi, however,had relatively lower CV values compared to SD values (Fig. 2A and B).This suggests that larger, elevated topographic features may contributemore to the surface variability in this floodplain than subtle topographicvariation at low elevations. In the Mississippi, Narran, and Macquarie,the opposite was observed, where CV values were relatively highercompared to the other floodplains and in comparison to what their SDvalues were (Fig. 2A and B). Relatively higher CV values indicate thatincreased surface variability is distributed at lower elevations in partsof these floodplains.
Skewness was highly variable in most floodplains and at mostwindow sizes (Fig. 2C). The majority of sites within the Shingwedzi,Bidgee, Macquarie, and Yanga floodplains had negatively skewed eleva-tion; the magnitude of which increased with window size. NegativeSkewness reflects the presence of a dominant elevation range in thesefloodplains uponwhich channels are incised. Themajority of sites with-in the Mississippi floodplain had positive Skewness at all window sizes,as did themajority of sites in theWoodforde floodplain at 50 and 200m.Positive Skewness indicates that these floodplain surfaces contained adominant height range upon which elevated features have beendeposited.
Differences in the four surfacemetrics resulted in statistically signif-icant differences among the eightfloodplains inmultidimensional spaceat the 50-m (PERMANOVA: pseudo-F= 12.08; d.f. = 7, 392; p=0.001;no. unique permutations = 998), 200-m (pseudo-F = 27.75; d.f. = 7,392; p = 0.001; no. perms. = 998), and 1000-m (pseudo-F = 97.17;d.f. = 7, 392; p=0.001; no. perms. = 996) window sizes. The numberof pairwise separations increased with window size, as did the strengthof separation (Table 2). At the 50-mwindow size, 15 of the 28 pairwisecomparisons were at least weakly separated (ANOSIM R N 0.25), with amean ANOSIM R value of 0.30 and a maximum of 0.70. At the 200-mwindow size, this increased to 18 of the 28 comparisons, with a meanof 0.36 and a maximum of 0.79. The sharpest increase in these figuresoccurred between the 200- and 1000-m window sizes; 27 of the 28comparisons were at least weakly separated, with a mean of 0.59and maximum of 0.94 between the Gwydir and Mississippi. This
Table 2Pairwise ANOSIM R values between each of the floodplains at each window size.
indicates that quantitative differences in the character and variabilitybetween these eight floodplain surfaces become more apparent, andmore readily detectable by these surface metrics, when measured acrosslarger areas.
4.2. Spatial organisation of surface character and variability
The spatial organisation of surface character differed among theeight floodplains and across sampling scales. From an almost ran-dom, unstructured distribution of metric values (e.g., models witha very low nugget; Fig. 3A); to highly fragmented patches of similarsurface character (e.g., models with a high nugget but low range;Fig. 3B); to relatively simple spatial gradients (e.g., models with ahigh range; Fig. 3C). An exponential isotropic model fit the spatialcorrelograms for all floodplains, metrics, and window sizes. Themodel nugget (strength of spatial organisation) and range (lengthscale of spatial organisation) parameters were highly variable anddependeduponfloodplain,metric, andwindowsize (Fig. 4). For example,theWoodforde had the three highest model nuggets overall, whichwere1.65 for CV at 200 m and 1.64 for CV at 50 m and SD at 200 m (Fig. 4).These high model nuggets indicate the presence of highly organisedpatches in these surface metric maps, associated with the main channel,distributaries, and splay formations (Fig. 3B). The Bidgee also had ahigh nugget for SD and CV at both the 50- and 200-m window sizes(Fig. 4). This reflected large areas of relatively low and constantvalues of these metrics, dissected by patches of high values associat-ed with meander cutoffs and scrolls. By comparison, the Gwydir hadthe three lowest model nuggets overall, which were 0.07 for Skew-ness at 50 m, 0.13 for SD at 50 m, and 0.14 for Skewness at 200 m.The Woodforde also had a very low nugget of 0.15 for Skewness at50 m (Fig. 4). These low model nuggets indicate poor spatial organi-sation in surface metric values. That is, areas of high and low valuesare randomly distributed throughout the floodplain, often resulting ina highly ‘grainy’ map of surface metric values (Fig. 3A). This is in con-trast to the highly patchy maps associated with high model nuggets(Fig. 3B).
Spatial model parameters depended largely upon the surface metricwithin some floodplains, even at the samewindow size. TheMississippiand the Narran had a nugget for CV more than 50% higher than that ofSD at the 50-m window size, whereas the other six floodplains hadsimilar nuggets between CV and SD (Fig. 4A and B). This indicates thatCV is more strongly organised in space than SD in these two floodplains,when measured at small window sizes. This is a result of higher surfacevariability distributed at low elevations throughout particular areas inthese two floodplains, which amplifies the strength of organisation ofCV values (Fig. 5).
The correlogram model nugget increased with window size in thefour unconfined floodplains for all metrics except CV in Narran andSkewness in Yanga, which were both lowest at 200 m. However, in thefour confined floodplains, the change in model nugget with windowsize was not consistent across metrics or floodplains. This indicatesthat, in the four unconfined floodplains, surfacemetric values are poorlyorganised when measured at smaller window sizes, but more highlyorganised at large window sizes. However, in the four confined flood-plains, the measurement scale at which floodplains have the strongestdegree of spatial organisation greatly depends upon the floodplainand the metric (Fig. 4).
The spatial correlogrammodel ranges were almost invariably higherin the four unconfined floodplains than in the four confined ones(Fig. 4). Some of the highest model ranges were observed in the Narranfor Skewness and SDCURV at all window sizes, as well as for SD at 50 m(Fig. 4). This indicates larger scale patchiness in these surface metricsthroughout theNarran. For Skewness, this spatial organisationwas asso-ciated with large areas of negative Skewness around the highlychannelised areas, positive Skewness around the banks of the lakes,and zero values in the flatter lake beds (Fig. 6A). For SDCURV, this was
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Nug
get (
Mor
an's
I)
0
2000
4000
6000
8000
10000
Ran
ge (
m) 50 m
200 m
1000 m
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Nug
get (
Mor
an's
I)
0
2000
4000
6000
8000
10000
Ran
ge (
m)
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Nug
get (
Mor
an's
I)
0
4000
8000
12000
16000
20000
Ran
ge (
m)
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Nug
get (
Mor
an's
I)
0
2000
4000
6000
8000
10000
Ran
ge (
m)
(D)
(C)
(B)
(A)
Fig. 4. Spatial correlogrammodel nugget (left) and range (right) for eachfloodplain at each sampling scale for (A) SD, (B) CV, (C) Skewness (note different range axis scale), and (D) SDCURV.Error bars indicate the standard error of the model parameter estimate.
109M.W. Scown et al. / Geomorphology 245 (2015) 102–116
associated with areas of highly dissected reticulate planform in thenortheastern part of the floodplain which created large patches of in-creased SDCURV (Fig. 6B). These areas also contributed to the highmodel range for SD at 50 m.
Themain exception to the rule of higher ranges in unconfined flood-plains was the Woodforde, whose ranges for SD and Skewness at the1000-m window size were two of the four highest overall. These largemodel ranges generally indicate a simpler, less fragmented spatial orga-nisation of these surfacemetrics in theWoodforde, which in some caseswas a simple gradient (Fig. 3C). TheMississippi, another confined flood-plain, also had a higher range than some unconfined floodplains for CVat the 50- and 200-m window sizes, as well as for SD at 50 m (Fig. 4Aand B). These inflated model ranges for CV in theMississippi were asso-ciated with large areas of high surface variability distributed at lowelevations in this floodplain, creating large patches of higher CV values(Fig. 5A).
The correlogram model range increased with window size for mostmetrics and floodplains (Fig. 4). This indicates that the distance overwhich surface metric values are spatially organised generally increaseswith the measurement window radius. However, a number of excep-tions to this occurred. A decrease in model range with increasingwindow size was observed for SD in Narran and for Skewness in Narranand Yanga. This was associated with large areas of highly channelisedreticulate planform in these floodplains, which led to surface metriccorrelations over large areas when measured at small window sizes,but shortened spatial autocorrelation when measured at large windowsizes (Figs. 6A and C). Other exceptions were the model ranges for theMississippi, which remained relatively constant over all window sizes(Fig. 4). This indicates that the characteristic length scale of organisationwithin the Mississippi was not influenced by window size across thescales investigated here, although the strength of organisation (modelnugget) was (Fig. 4).
Fig. 5. Example of differences in the strength of spatial organisation at the 50-mwindow size between (A) CV and (B) SD inMississippi; and between (C) CV and (D) SD in Narran. Spatialcorrelograms and models are shown below with Moran's I from −1 to 1 on the y-axis and lag distance on the x-axis.
110 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
4.3. Associations between environmental variables and components ofsurface complexity
Floodplain dissimilarities based on surface metrics were most corre-lated with differences in floodplain discharge variables and sedimentyield at the smaller window sizes and with differences based solely onsediment yield at the largest window size (Table 3). The maximal corre-lation coefficient increasedwithwindow size from 0.35 to 0.55 (Table 3).At the 50-m window size, greatest correlation was associated with thefour discharge variables and sediment yield. At 200 m, the optimumcombination reduced to three discharge variables (Q, QCV, QCVAnn) andsediment yield. At 1000 m, sediment yield was the sole environmentalvariable responsible for a maximal correlation of 0.55 with differencesin floodplain surface metrics.
Floodplain dissimilarities based on spatial model parameters weremost correlated with differences in floodplain width at the smallerwindow sizes andwith differences in valley slope at the largest windowsize (Table 3). Correlationwas highest at the 200-mwindow size, with acoefficient of 0.65 associated solely with floodplain width. Floodplainwidth was also the sole environmental variable responsible for a maxi-mal correlation of 0.38with differences in the spatial organisation of thefloodplain surfaces at the 50-m window size. At 1000 m, valley slope
was the sole environmental variable associated with floodplain spatialmodel dissimilarities, albeit with a low correlation coefficient of 0.14.
The results of the bioenv analysis were supported by environmentalvector correlations with the surface metric ordinations of floodplains(Fig. 7A, B, and C; Table 4). Again, the highest correlations were associ-ated with discharge variables and sediment yield for the surface metricordination, although floodplain width and valley slope were also signif-icant (Table 4). In particular, Mississippi had very high Q relative to allother floodplains, and Mississippi sites were separated from otherfloodplains along this axis, particularly at the 1000-m window size(Fig. 7A, B, and C). The Q vector direction was between those axes ofCV and Skewness, whichwere correlatedwith the surfacemetric ordina-tion patterns (Table 5), indicating thatQmaybe related to differences inthese two metrics across the eight floodplains. Significant correlationsalso occurred between QCV, QCVAnn, and QCVMax and the surface metricordinations along a similar vector axis to that of Skewness. Two ofthese discharge variables were highest in Shingwedzi, and thesevectors were associated with the direction of separation of Shingwedzisites from other floodplains, particularly at the 1000-m window size(Fig. 7C). Shingwedzi also had the highest SY and the lowest Fpw, andthese variable vectors were oriented along similar axes to those of SDand SDCURV. Floodplain width was also associated with the clumping of
Fig. 6. Example of differences in spatial organisation between (A) Skewness at the 50-mwindow size; (B) SDCURV at the 200-mwindow size; and (C) Skewness at the 1000-mwindow sizefor Narran. Spatial correlograms and models are shown below with Moran's I from −1 to 1 on the y-axis and lag distance on the x-axis.
111M.W. Scown et al. / Geomorphology 245 (2015) 102–116
Narran, Yanga, Gwydir, and Macquarie sites in the surface metric ordi-nations (Fig. 7A, B, and C). This vector direction was almost the directopposite of that of SD at all window sizes. Althoughmost of the environ-mental variables were significantly correlated with the surface metricordination patterns, they had low r2 values (Table 4). This suggeststhat factors other than the large-scale environmental conditions within
Table 3Associations between floodplain dissimilarity structures based on surface metrics and onspatial organisation model parameters at each window size from bioenv analysis.
Optimum combinationof variables
Correlationcoefficient
Surface metrics 50 m Q, QCV, QCVAnn, QCVMax, SY 0.35200 m Q, QCV, QCVAnn, SY 0.40
1000 m SY 0.55Spatial model parameters 50 m Fpw 0.38
200 m Fpw 0.651000 m Vs 0.14
which these floodplains occur are causing high variability of surfacecharacter among sites within a single floodplain. All surface metricswere strongly and significantly associatedwith themultivariate ordina-tion patterns among floodplain sites (Table 5).
Three groups of floodplains emerged at each window size based onthe eight spatial organisation variables (i.e., the nugget and range foreach of the four surface metrics). This separation of groups of flood-plains was most strongly correlated with Fpw across all window sizes(Table 4), although the spatial model parameters associated with theseparations differed (Table 5). At the 50-mwindow size, the Narran clus-tered alone; while the three most confined floodplains (the Shingwedzi,Woodforde, and Bidgee) were also separated from the remaining fourfloodplains (Fig. 7D). The separation of floodplain spatial organisationwas strongly and significantly associated with the model ranges for SDand Skewness (Fig. 7D; Table 5); however, no environmental variableswere consistent with the influence of these spatial model parameters atthis scale. At the 200-m window size, the Gwydir clustered alone; whilethe Shingwedzi, Woodforde, and Bidgee were again separated fromthe other floodplains. This separation was strongly and significantly
Fig. 7. Associations between environmental variables, surface metrics, and floodplain ordinations at window sizes of (A) 50-m radius, (B) 200-m radius, and (C) 1000-m radius; andbetween environmental variables, spatial model parameters, and floodplain spatial organisation ordinations at window sizes of (D) 50-m radius, (E) 200-m radius, and (F) 1000-m radius.Vectors show direction of significant (p b 0.05) associations between environmental variables and ordination patterns. *Indicates not significant at p= 0.05.
112 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
associated with the model ranges for SD, CV, and SDCURV, as well as thenuggets for SD and CV (Fig. 7E; Table 5). The direction of the Fpw vectorwas consistent with those of SD and CV ranges (Fig. 7E), indicating thatFpw may be particularly influential in structuring these spatial charac-teristics of floodplain surfaces. In particular, wider floodplains such asthe Macquarie and Gwydir also appear to be spatially organised overlonger distances (i.e., larger patches) in terms of their surface SD andCV. At the 1000-m window size, the Woodforde clustered alone, whilethe Mississippi and Narran were also separated from the remainingfive floodplains (Fig. 7F). Within the group of five floodplains, the
Table 4Correlations between environmental variables and the ordination patterns based on(a) surfacemetrics and (b) spatial model parameters (* indicates not significant at p=0.05;bold indicates vectors shown on ordination plot; p and r2 values are provided).
Shingwedzi, Bidgee, and Yangawere tightly clumped inmultidimension-al space, as were the Macquarie and Gwydir. Separations of floodplainspatial organisation were again highly correlated with Fpw (Fig. 7F;Table 4), which also appeared to be associated with separations along asimilar vector axis to that of the model range for CV. These results againsupported the results of the bioenv analysis at the two smaller windowsizes. However, at the 1000-m window size, Vswas identified as the en-vironmental variable most correlated with spatial organisation dissimi-larities in the bioenv analysis (Table 3), while Fpw was the variablemost associated with ordination patterns (Fig. 7F).
Table 5Correlations between ordination patterns and variables for (a) surface metrics and(b) spatial model parameters at each window size (* indicates not significant at p = 0.05;bold indicates significant vectors shown on ordination plot; p and r2 values are provided).
113M.W. Scown et al. / Geomorphology 245 (2015) 102–116
5. Discussion
5.1. Floodplain surface character and variability
Our results show that surface character and variability is highlyvariable within and among floodplains, and these differences can bequantified using a suite of surface metrics (sensu McGarigal et al.,2009). The four surface metrics used in this study distinguished thespatial complexity between eight floodplain surfaces (Table 6). Somefloodplains were relatively flat, with low relief, but with subtle topo-graphic variability (e.g., Gwydir andNarran). Otherfloodplainshadhigherrelief associatedwith large, distinct surface features (e.g., Shingwedzi andBidgee). Floodplain physical features that contributed to differences inspatial complexity varied from the presence of a single, deeply incisedmeandering channel that dissected the floodplain (e.g., Shingwedzi)to those with large areas of highly channelised reticulate planforms(e.g., Macquarie). Other significant physical features indicated by thesurface metrics included natural levees, scroll swales, and crevassesplays. The presence and spatial organisation of these different physicalfeatures across floodplain surfaces contribute to their overall spatialcomplexity (Baker and Barnes, 1998; Richards et al., 2002; Hamiltonet al., 2007).
The four surface metrics used in this study were able to distinguishparticular surface attributes throughout the eightfloodplains. An example
Table 6General distinguishing attributes of the spatial complexity of the eight floodplain surfaces.
Floodplain Surface character and variability Spatial organisationof surface conditions
Shingwedzi High magnitude of topographicrelief and variability associatedwith deeply incised mainchannel
Very strongly organised overshort distances owing to distinctsurface features such as steepchannel banks clustered oversmall spatial extents
Woodforde High topographic variabilityacross small areas associatedwith main channel, naturallevees, and crevasses splays
Highly patchy in some conditionsassociated with main channeland splay formations, but otherconditions organised along asimple longitudinal gradient
Bidgee Increased topographic variabilityaround deep meander cutoffsand scroll swales
Very strongly organised overshort distances owing to distinctsurface features such as meandercutoffs clustered over smallspatial extents periodicallydownstream
Mississippi Increased topographic variabilityat low elevations associated withbackswamps and islands
Spatially organised over very largescales owing to distinct surfaceconditions over large areas
Yanga High variability of erosional anddepositional areas andassociated complex channelnetworks and variablemorphological forms
Highly fragmented landscape atsmall sampling scales owing tohighly variable geomorphicconditions, tending to moderateorganisation at large scales
Macquarie Low magnitude of relief buthighly convoluted surfaceassociated with channelisedreticulate planform throughout
Large-scale organisation ofvariability at low elevationsassociated with large areas ofreticulate channels; absence ofpatches associated with largesurface features
Gwydir Relatively flat and featurelessover large areas
Generally weak spatialorganisation at small samplingscales owing to lack of smalldistinct features, tending tolarge-scale patchiness at largesampling scales owing topersistence of surface conditionsover large areas
of this was the ability of CV and SDCURV to distinguish highly channelisedreticulate areas, particularly in Narran andMacquarie. Reticulate channelmorphology is common in arid and semiarid floodplains throughoutAustralia (Fagan and Nanson, 2004). Our results suggest that by mea-suring these metrics in a moving window across floodplain DEMs,quantitatively distinguishing and potentially delineating areas of retic-ulate channel morphology is possible. Other surface metrics are alsolikely able to determine particular fluvial morphologies, which tradi-tionally has been a qualitative assessment (Nanson and Croke, 1992;Fagan and Nanson, 2004; Lewin and Ashworth, 2014).
The four surfacemetrics used in this studymeasure different aspectsof floodplain surface character and variability. The SD of surface heightsreflects themagnitude of topographic relief and its variability throughouta floodplain surface. This is important for providing a range of floodfrequencies and durations, soil moisture conditions, soil texture andorganic matter concentrations, plant species communities, and nutrientdynamics throughout a floodplain (Medley, 1992; Pollock et al., 1998;Pinay et al., 2000; De Jager et al., 2012). It is also likely to exert morecontrol over spatial inundation patterns because larger, steeper surfacefeatures have an enhanced ability to direct flow across the floodplain(Alsdorf et al., 2007). The CV of surface heights distinguishes surfaces inwhich greater topographic variability is distributed in lower topographicareas. This may be particularly important given the disproportionatelylarge ecological effects that small, frequent floods can have on the struc-ture and function offloodplain ecosystems (Buchholz, 1981; Pollock et al.,1998; Hamilton et al., 2007). Thus, small changes in water surface eleva-tion during small floods, or at the beginning and end of larger floods,could have large ecological effects in floodplains with high CV becauseof increased surface variability at low elevations. For example, Pollocket al. (1998) found the coefficient of variation of flood frequency (basedon surface elevation) to have a significant effect on plant species richnessin Alaskan floodplains. Skewness indicates whether a surface has a domi-nant height range that the surface has been reduced below or increasedabove. This measure can be used to infer if floodplain surfaces havebeen altered predominantly via erosion or aggradation and thuswhethererosion or deposition is dominating the physical structure of a floodplainsurface, or parts thereof. Both highly positive and highly negativeSkewness values were observed among the eight floodplains studied,indicating different structuring processes dominate these floodplainsurfaces. The results of this study suggest that the Shingwedzi and Bidgeeare examples of predominantly eroding or incising floodplain surfaces(negative Skewness), while the Woodforde and Mississippi are exam-ples of aggrading floodplain surfaces (positive Skewness). However, awide range of Skewness values were also observed within floodplains,particularly in the Bidgee, Narran, and Yanga. This supports the findingsof Asselman and Middelkoop (1995), Walling and He (1998), andSouthwell (2008) that erosion and deposition within floodplains can behighly variable in space. Surface roughness or convolutedness is indicatedby SDCURV. However, SDCURVwasmeasured here as the standard deviationof total curvature, which is sensitive to horizontal and to vertical variabil-ity in the surface. The standard deviation of only plan curvature (sensuJenness, 2012) may be more appropriate in measuring floodplain sur-faces, as thiswould emphasise lateral surface variability and be less sensi-tive to vertical relief, which is alreadymeasured directly using SD and CV.
5.2. Spatial organisation of surface conditions
The spatial organisation of floodplain surface conditions was highlyvariable among those studied, and both the strength and scale of thisorganisation distinguishes between groups of floodplains (Table 6).The differences between floodplains reflect the distinct spatial arrange-ments of surface features. Some floodplains contained small areas ofhighly similar surface conditions, which were in distinct contrast toneighbouring areas; whereas conditions in other floodplains changed ina continuous and gradual manner either downstream or laterally acrossthe surface. This confirms that floodplain landscape patterns may exist
114 M.W. Scown et al. / Geomorphology 245 (2015) 102–116
as a patch mosaic or as a gradient, depending upon the floodplain(Naiman et al., 1988; van Coller et al., 2000; Wiens, 2002).
The spatial organisation of landscape patterns can indicate thenature of structuring processes and can have ecological consequences(Turner and Gardner, 1990; Holling, 1992; De Jager and Rohweder,2012). Patchiness in floodplain surface conditions might indicate thatpositive hydrogeomorphic feedback loops exist, which reinforce a par-ticular surfacemorphology once it has been established. Such feedbacksare known to be associated with vegetation in island-dominated chan-nels and floodplains (Gurnell and Petts, 2002). Gradients, on the otherhand, might indicate that distance downstream or from the channelcontrols how surface morphology varies in space throughout a flood-plain, affecting patterns either longitudinally or laterally. Other spatialpatterns in surface conditions observed in this study appeared random,with little to no spatial autocorrelation. This might suggest that thereare no feedbacks between surface morphology and hydrogeomorphicprocesses and that conditions occur randomly or,more likely, that orga-nisation exists at a scale smaller than those investigated in this study.
5.3. Associations with environmental variables
Our results indicate that discharge, energy conditions, and sedimentwere associated statistically with different aspects of the spatialcomplexity of floodplain surfaces. Sediment yield and discharge wereprimarily associated with the character and variability of elevationthroughoutfloodplain surfaces,whereas energy conditionswere associ-ated with the spatial organisation of the surface. Sediment yield wasexpected to be highly associated with floodplain surface complexity atall scales (Church, 2002). In terms of floodplain surface character andvariability, our results support this, although discharge variables alsoappear to be important at smaller sampling scales. In particular, increas-ing sediment yield and flow variability was associated with a greatervariability of topographic relief, while mean discharge was associatedwith increased variability at low elevations. Others have suggestedthat different valley trough environments lead to quite different flood-plain morphologies occurring within them (Hynes, 1975; Warner,1992). In terms of the spatial organisation of surface conditions, our re-sults support this, with floodplain width having the greatest associationwith differences in spatial organisation among the floodplain sur-faces at two of the three sampling scales andwith valley slope havingso at the largest sampling scale, although the relationship was weak.In particular, increasing floodplain width was associated with largerscale organisation in surface conditions.
Significant associations between environmental variables and thespatial complexity of the eight floodplain surfaces were observed inthis study, although correlation coefficients were low at some samplingscales. The weakness of these coefficients has a number of potentialcauses. First, the only sediment variablemeasured in this studywas sed-iment yield. However, sediment character is also likely to influencefloodplain morphology (Nanson and Croke, 1992), and inclusion ofsuch data into this study may substantially increase the correlationcoefficients in Table 4. However, such data were not available for all ofthe floodplains studied. Second, present floodplain morphologymay be influenced by shifts in past discharge and sediment regimes(Warner, 1992).We have onlymeasured environmental variables relat-ing to the present discharge and sediment regimes, whichmay not fullyaccount for differences in the spatial complexity of surfaces among theeight floodplains. Third, vegetation is known to be highly influential infloodplain biogeomorphology (Nanson and Beach, 1977; Salo et al.,1986; Hupp andOsterkamp, 1996; Hughes, 1997). Inclusion of informa-tion on vegetation types, life strategies, and structural characteristics inthese floodplains may have contributed to more complete links be-tween the environment and spatial complexity of floodplain surfaces.Finally, a scale mismatch in the measurement of environmental vari-ables and surface attributes may have occurred. Surface conditionswere highly variable within each floodplain; however, only a single
value of each environmental variable was attributed to each floodplain.This is likely the cause of significant but low correlation coefficients inTable 4. Accounting for the internal variability of environmental vari-ables within floodplains may be more appropriate, particularly forflow and sediment delivery, as these may vary greatly throughout asingle floodplain (Asselman and Middelkoop, 1995; Walling and He,1998). If significant associations between surface complexity and flowand sediment conditions within floodplains could be established, mea-surement of floodplain surface complexity using LiDAR may provide auseful surrogate to determine the spatial variability of hydrogeomor-phic conditions across floodplain surfaces.
5.4. Scale
Floodplains are influenced by geomorphic factors operating acrossmultiple scales (Hughes, 1997; Notebaert and Piégay, 2013). The spatialand temporal scale at which causal factors operate is related to thescales over which they influence landscape patterns (Schumm andLichty, 1965). Our results indicate that different external factors influ-ence floodplain surface complexity at different spatial scales, and thusoperate over different scales in floodplains. Sampling scale influencedthe two components of floodplain surface complexity and their associa-tions with environmental variables. Across the eight floodplains stud-ied, our results indicate that sediment yield may be an importantdriver of surface character and variability across a range of scales(101–103 m), while flow discharge and variability may be influentialat smaller scales of 101–102 m within floodplains. The influence offloodplain width on the spatial organisation of surface conditions ap-pears to be relatively consistent across all sampling scales investigatedin these eightfloodplains. This suggests that floodplainwidth influencessurface complexity from small spatial scales within the floodplain to thescale of the entire floodplain itself.
This study examined the effect of scale on surface complexity and itsassociation with environmental drivers within floodplains; however,floodplains are also nested at a particular scale within the river systemhierarchy (Frissell et al., 1986; Dollar et al., 2007). Different levelswithinthis hierarchy are also known to be structured by different externalfactors; for example, processes structuring a river reach are vastly differ-ent and operate at different spatial and temporal scales from thoseprocesses structuring entire stream networks (Phillips, 1988; Thorpet al., 2008). Structural patterns are also known to self-emerge fromsmall-scale interactions within geomorphic systems (Hallet, 1990), asis typical of complex systems in general (Phillips, 1999). Quantitativedata such as LiDAR will likely prove useful in investigating such phe-nomena in floodplains in the future.
Sampling scale also affected how quantitatively different floodplainsurfaces are when measured at different scales. In terms of the surfacecharacter and variability, sampling scales of at least 1000-m radiusenabled these four surface metrics to more readily detect quantitativedifferences between floodplain surfaces and their associations with en-vironmental variables. This was evidenced by the fact that floodplainsbecamemore distinguishable from each other at larger sampling scales,and associations with environmental variables became stronger andmore clearly related to sediment yield at larger sampling scales. However,sampling scales N 1000 m have rarely been adopted in floodplain land-scape research (Scown et al., in press) and quantification of floodplainlandscape structure may need to be targeted at larger scales than hasbeen the tradition. The spatial organisation of surface conditions indicatea less consistent effect of sampling scale. In the four unconfined flood-plains, a general trend from weak organisation at small sampling scalesto strong organisation at large sampling scales occurred. This increasein strength of organisation is consistent with the ‘filtering’ of fine-grained spatial patterns, which occurs when using moving windows (orobservational areas) of increasing size. However, in the confined flood-plains, no consistent trend in the strength of organisationwith increasingsampling scale was observed. This indicates that, in confined floodplains,
115M.W. Scown et al. / Geomorphology 245 (2015) 102–116
the sampling scale at which surface conditions are most stronglyorganised depends on the floodplain. In general, the length scale of or-ganisation also increasedwith sampling scale, again consistent with the‘filtering’ of spatial patternswith increasing observation scale; however,Narran was an exception to this.
6. Conclusions
Quantitative investigations of floodplain surface complexity atdifferent scales and in different geomorphic settings are required to de-velopmore general theories of the linkages between landscape patternsand larger-scale driving variables. The increasing availability of high res-olution spatial data over large areas makes this task more achievablethan ever. Our results indicate that, across the eight floodplains studied,sediment yield is associated with differences in surface character andvariability across a range of scales (101–103 m), while flow dischargeand variability are also associated with surface morphology at smallersampling scales (101–102 m). Our results also indicate that floodplainwidth is highly associatedwith the spatial organisation of surface condi-tions, particularly around the 102–103m sampling scales in these flood-plains. These results confirm that the environmental variables primarilydriving floodplain patterns likely depend upon the scale at which thosepatterns are observed. Particular environmental variables also appearedto be related to ordination separation of floodplains in the same direc-tions as particular surface characteristics, indicating that environmentalvariables may differentially influence surface attributes. In particular,discharge variables were associated with floodplain separation alongaxes similar to CV and Skewness; sediment yield was associated withseparations along a similar axis to SD and SDCURV; and floodplain widthwas associated with separations related to SD and SDCURV, but also tothe length scale of organisation of SD and CV. Differences in surface char-acter and variability among the eight floodplains also increased withsampling size, as did the strength of their associationwith environmentalvariables, suggesting that small-scale studies may overlook landscape-scale differences between floodplains. Although eight floodplains is asmall sample size, it is an advancement on traditional, often qualitative,single-floodplain investigations of this nature. As LiDAR data for morefloodplain surfaces become available to researchers, further investiga-tions should be conducted to evaluate whether the general patterns ob-served in this study transfer to other vastly different geographic settings.
Acknowledgements
The authors wish to thank Richard Marston and two anonymousreviewers, whose input has greatly improved this manuscript. We alsoacknowledge support from the University of New England and theUSGS Upper Midwest Environmental Sciences Center, without whichthis research would not have been possible. Any use of trade, product,or firm names is for descriptive purposes only and does not implyendorsement by the U.S. Government.
References
Alsdorf, D., Bates, P., Melack, J., Wilson, M., Dunne, T., 2007. Spatial and temporal complexityof the Amazon flood measured from space. Geophys. Res. Lett. 34, L08402.
Anderson, M.J., 2001. A newmethod for non-parametric multivariate analysis of variance.Austral Ecol. 26, 32–46.
Baker, M.E., Barnes, M.V., 1998. Landscape ecosystem diversity of river floodplains innorthwestern Lower Michigan, U.S.A. Can. J. For. Res. 28, 1405–1418.
Bar Massada, A., Radeloff, V.C., 2010. Two multi-scale contextual approaches for mappingspatial pattern. Landsc. Ecol. 25, 711–725.
Buchholz, K., 1981. Effects of minor drainages on woody species distributions in a succes-sional floodplain forest. Can. J. For. Res. 11, 671–676.
Church, M., 2002. Geomorphic thresholds in riverine landscapes. Freshw. Biol. 47,541–557.
Church, M., Mark, D.M., 1980. On size and scale in geomorphology. Prog. Phys. Geogr. 4,342–390.
Clarke, K.R., 1993. Non-parametric multivariate analyses of changes in community structure.Aust. J. Ecol. 18, 117–143.
Clarke, K.R., Ainsworth, M., 1993. A method of linking multivariate community structureto environmental variables. Mar. Ecol. Prog. Ser. 92, 205–219.
De Jager, N.R., Rohweder, J.J., 2012. Spatial patterns of aquatic habitat richness in theUpper Mississippi River floodplain, USA. Ecol. Indic. 13, 275–283.
De Jager, N.R., Thomsen, M., Yin, Y., 2012. Threshold effects of flood duration on thevegetation and soils of the Upper Mississippi River floodplain, USA. For. Ecol.Manag. 270, 135–146.
De Rose, R.C., Prosser, I.P., Wiesse, M., Hughes, A.O., 2003. Patterns of erosion and sedimentand nutrient transport in the Murray–Darling Basin. Technical Report 32/03, CSIROLand and Water, Canberra.
Dollar, E.S.J., James, C.S., Rogers, K.H., Thoms,M.C., 2007. A framework for interdisciplinaryunderstanding of rivers as ecosystems. Geomorphology 89, 147–162.
Fagan, S.D., Nanson, G.C., 2004. The morphology and formation of floodplain-surfacechannels, Cooper Creek, Australia. Geomorphology 60, 107–126.
Frissell, C.A., Liss, W.J., Warren, C.E., Hurley, M.D., 1986. A hierarchical framework forstream habitat classification: viewing streams in a watershed context. Environ.Manag. 10, 199–214.
Gurnell, A.M., Petts, G.E., 2002. Island-dominated landscapes of large floodplain rivers, aEuropean perspective. Freshw. Biol. 47, 581–600.
Hallet, B., 1990. Spatial self-organization in geomorphology: from periodic bedforms andpatterned ground to scale-invariant topography. Earth Sci. Rev. 29, 57–75.
Hamilton, S.K., Kellndorfer, J., Lehner, B., Tobler, M., 2007. Remote sensing of floodplaingeomorphology as a surrogate for biodiversity in a tropical river system (Madre deDios, Peru). Geomorphology 89, 23–38.
Heylighen, F., 1989. Self-organisation, emergence and the architecture of complexity.Proceedings of the 1st European Conference on System Science, Paris, France.
Holling, C.S., 1992. Cross-scale morphology, geometry, and dynamics of ecosystems. Ecol.Monogr. 62, 447–502.
Hudson, P., Butzer, K., Beach, T., 2008. Fluvial deposits and environmental history: synthesis.Geomorphology 101, xvii–xx.
Geomorphology 14, 277–295.Hynes, H., 1975. The stream and its valley. Verh. Int. Ver. Limnol. 19, 1–15.Ibisate, A., Ollero, A., Díaz, E., 2011. Influence of catchment processes on fluvial morphology
and river habitats. Limnetica 30, 169–182.Jenness, J.S., 2012. DEM Surface Tools. Jenness Enterprises ([Online] http://www.
jennessent.com/arcgis/surface_area.htm).Kotliar, N.B., Wiens, J.A., 1990. Multiple scales of patchiness and patch structure: a hierar-
chical framework for the study of heterogeneity. Oikos 59, 253–260.Legleiter, C.J., 2014. A geostatistical framework for quantifying the reach-scale spatial
structure of river morphology: 1. Variogram models, related metrics, and relationto channel form. Geomorphology 205, 65–84.
Lewin, J., Ashworth, P.J., 2014. Defining large river channel patterns: alluvial exchangeand plurality. Geomorphology 215, 83–98.
Lvovich, M.I., Karasik, G.Y., Bratseva, N.L., Medvedeva, G.P., Maleshko, A.V., 1991. Contempo-rary Intensity of the World Land Intracontinental Erosion. USSR Academy of Sciences,Moscow.
McGarigal, K., Cushman, S., Stafford, S., 2000. Multivariate Statistics for Wildlife and EcologyResearch. Springer-Verlag, New York.
McGarigal, K., Tagil, S., Cushman, S., 2009. Surface metrics: an alternative to patch metricsfor the quantification of landscape structure. Landsc. Ecol. 24, 433–450.
Medley, K.E., 1992. Patterns of forest diversity along the Tana River, Kenya. J. Trop. Ecol. 8,353–371.
Msadala, V., Gibson, L., Le Roux, J., Rooseboom, A., Basson, G.R., 2010. Sediment Yield Predic-tion for South Africa: 2010 Edition. South AfricanWater Research Commission, Pretoria.
Naiman, R.J., Décamps, H., Pastor, J., Johnston, C.A., 1988. The potential importance ofboundaries of fluvial ecosystems. J. N. Am. Benthol. Soc. 7, 289–306.
Nanson, G.C., Beach, H.F., 1977. Forest succession and sedimentation on a meandering-river floodplain, Northeast British Columbia, Canada. J. Biogeogr. 4, 229–251.
Nanson, G.C., Croke, J.C., 1992. A genetic classification of floodplains. Geomorphology 4,459–486.
Notebaert, B., Piégay, H., 2013. Multi-scale factors controlling the pattern of floodplainwidth at a network scale: the case of the Rhône basin, France. Geomorphology 200,155–171.
Phillips, J.D., 1988. The role of spatial scale in geomorphic systems. Geogr. Anal. 20,308–317.
Phillips, J.D., 2007. Perfection and complexity in the lower Brazos River. Geomorphology91, 364–377.
Pinay, G., Black, V.J., Planty-Tabacchi, A.M., Gumiero, B., Décamps, H., 2000. Geomor-phic control of denitrification in large river floodplain soils. Biogeochemistry 50,163–182.
Pollock, M.M., Naiman, R.J., Hanley, T.A., 1998. Plant species richness in riparianwetlands — a test of biodiversity theory. Ecology 79, 94–105.
Radeloff, V.C., Miller, T.F., He, H.S., Mladenoff, D.J., 2000. Periodicity in spatial data andgeostatistical models: autocorrelation between patches. Ecography 23, 81–91.
Ren, Y.,Wei, X.,Wang, D., Luo, Y., Song, X.,Wang, Y., Yang, Y., Hua, L., 2013. Linking landscapepatterns with ecological functions: a case study examining the interaction betweenlandscape heterogeneity and carbon stock of urban forests in Xiamen, China. For. Ecol.Manag. 293, 122–131.
Richards, K., Brasington, J., Hughes, F., 2002. Geomorphic dynamics of floodplains: ecolog-ical implications and a potential modelling strategy. Freshw. Biol. 47, 559–579.
Salo, J., Kalliola, R., Häkkinen, I., Mäkinen, Y., Puhakka, M., Coley, P.D., 1986. River dynamicsand the diversity of Amazon lowland forest. Nature 322, 254–258.
Schumm, S.A., Lichty, R.W., 1965. Time, space, and causality in geomorphology. Am. J. Sci.263, 110–119.
Scown, M., Thoms, M., De Jager, N.R., 2015. Measuring spatial pattern in floodplains: astep towards understanding the complexity of floodplain ecosystems. In: Gilvear,D.J., Greenwood, M., Thoms, M., Wood, P. (Eds.), River Science: Research and Man-agement for the 21st Century. John Wiley & Sons, Chichester (in press).
Simon, H.A., 1962. The architecture of complexity. Proc. Am. Philos. Soc. 106, 467–482.Sokal, R.R., 1979. Ecological parameters inferred from spatial correlograms. In: Patil, G.P.,
Rosenzweig, M.L. (Eds.), Contemporary Quantitative Ecology and Related Ecometrics.Statistical Ecology Series vol. 12. International Co-operative Publishing House,Fairland, MD, pp. 167–196.
Southwell, M., 2008. Floodplains as DynamicMosaics: Sediment and Nutrient Patches in aLarge Lowland Riverine Landscape. University of Canberra, Canberra (Thesis submittedfor the degree of Doctor of Philosophy).
Thoms, M.C., 2003. Floodplain–river ecosystems: lateral connections and the implicationsof human interference. Geomorphology 56, 335–349.
Thoms, M.C., Parker, C.R., Simons, M., 2000. The dispersal and storage of trace metals inthe Hawkesbury River Valley. In: Brizga, S., Finlayson, B. (Eds.), River Management:The Australasian Experience. John Wiley & Sons, Chichester, pp. 197–219.
Thorp, J.H., Thoms, M.C., Delong, M.D., 2008. The Riverine Ecosystem Synthesis: TowardsConceptual Cohesiveness in River Science. Elsevier, Amsterdam.
Turner, M.G., Gardner, R.H. (Eds.), 1990. Quantitative Methods in Landscape Ecology: TheAnalysis and Interpretation of Landscape Heterogeneity. Springer-Verlag, New York.
Turner, R.E., Rabalais, N.N., 2004. Suspended sediment, C, N, P, and Si yields from theMississippi River Basin. Hydrobiologia 511, 79–89.
van Coller, A.L., Rogers, K.H., Heritage, G.L., 2000. Riparian vegetation-environment rela-tionships: complimentarity of gradients versus patch hierarchy approaches. J. Veg.Sci. 11, 337–350.
Walling, D.E., He, Q., 1998. The spatial variability of overbank sedimentation on riverfloodplains. Geomorphology 24, 209–223.
Warner, R.F., 1992. Floodplain evolution in a New South Wales coastal valley, Australia:spatial process variations. Geomorphology 4, 447–458.
Wiens, J.A., 2002. Riverine landscapes: taking landscape ecology into the water. Freshw.Biol. 47, 501–515.
Wu, J., 2004. Effects of changing scale on landscape pattern analysis: scaling relations.Landsc. Ecol. 19, 125–138.
Wu, J., Qi, Y., 2000. Dealing with scale in landscape analysis: an overview. Geogr. Inf. Sci. 6,1–5.
