Drainage Network Detection and Assessment of Network Storage Capacity in Agrarian Landscape

HYDROLOGICAL PROCESSESHydrol. Process. 27, 541–553 (2013)Published online 20 March 2012 in Wiley Online Library(wileyonlinelibrary.com) DOI: 10.1002/hyp.9224

Drainage network detection and assessment of network storagecapacity in agrarian landscape

Federico Cazorzi,1* Giancarlo Dalla Fontana,2 Alberto De Luca,3 Giulia Sofia2 and Paolo Tarolli21 Department of Agriculture and Environmental Science, University of Udine, via delle Scienze 208, 33100 Udine, Italy

2 Department of Land and Agroforest Environments, University of Padova, Agripolis, viale dell’Università 16, 35020 Legnaro (PD), Italy3 Geomatica & Ambiente s.r.l., viale dell’Università 14, 35020 Legnaro, Italy

*CEnvUdE-m

Co

Abstract:

Drainage networks in agrarian landscape within floodplains constitute surface’s discontinuities that are expected to affecthydrological response during floods. Drainage network recognition and quantification of water storage capacity within channelsare, therefore, crucial for watershed planning and management. These evaluations require accurate spatial information for thearea of interest and in most cases, when studying large catchments, broad datasets of ditches locations and descriptions are notavailable. In order to characterize drainage networks for large areas, the availability of high resolution topography derived byairborne laser scanner (LiDAR) represents a new and effective tool. Nowadays LiDAR DTMs covering large areas are readilyavailable for public authorities, and there is a greater and more widespread interest in the application of such information for thedevelopment of automated methods aimed at solving geomorphological and hydrological problems. While LiDAR DTMsreliability in steep landscape has been proven by several recent studies, only few researches have been conducted to take intoaccount the effectiveness of these data in agrarian low relief landscapes. The goal of this research is to propose a semi-automaticapproach based on a LiDAR DTM to (1) detect drainage networks in agrarian/floodplain contexts, and (2) to estimate some of thenetwork summary statistics (network length, width, drainage density and storage capacity). The procedure is applied in twotypical alluvial-plain areas in the North East of Italy, and tested comparing automatically derived network with surveyed ones.The results underline the capability of high resolution DTMs for drainage network detection and characterization in the contextof agrarian landscapes within floodplains, opening at the same time new challenges to evaluate some hydrological processes inthese areas. Copyright © 2012 John Wiley & Sons, Ltd.

KEY WORDS artificial drainage network; agrarian landscape; high resolution topography; DTM; LiDAR

Received 5 August 2011; Accepted 13 January 2012

INTRODUCTION

Drainage networks in agrarian landscape within flood-plains constitute man-made surface’s discontinuities, andthey are expected to affect hydrological response duringflood events. Drainage networks, in fact, influencegroundwater hydrology (Dagès et al., 2009), but alsooverland flow paths (Duke et al., 2006; Gascuel-Odouxet al., 2011; Levavasseur et al., 2011). Consideringagricultural ditch drainage networks, four effects arecommonly accounted for: the interception of overlandflow on hillslopes, the drainage of groundwater bylowering the water table, infiltration from the ditchtowards the groundwater, and conveyance of watertowards downstream areas (Carluer and Marsily, 2004;Dunn and Mackay, 1996; Levavasseur et al., 2011).Clearly, these networks influence the water transfer(Moussa et al., 2002), and they play a key role in thecontrol of floods generation (Gallart et al., 1994; Voltzet al., 1998; Marofi, 1999; Moussa et al., 2002).

orrespondence to: Federico Cazorzi, Department of Agriculture andironmental Science, University of Udine, via delle Scienze 208, 33100ine, Italy.ail: [email protected]

pyright © 2012 John Wiley & Sons, Ltd.

In Northern Italy, and in particular in alluvial plains,massive investments in land reclamation have alwaysplayed an important role for flood protection. Theextensive drainage network designed for reclamationpurposes was able, in the past, to control groundwaterlevel, especially when dealing with long-term heavyrainfall, which are the most critical in these areas. Duringthe last years, changes in agricultural techniques andthe simultaneous increase of urbanization andindustrialization increased the extent of impervious areasand reduced the extent of agrarian drainage networks. Asa consequence, an overall decrease in volumes of smallreservoirs is generally registered, and some studieshighlighted the consequent effect on flood risks (Brathet al., 2003), pointing out the impacts of urbanization andchanging agricultural practices in flood flows. These landuse changes, on the one hand, greatly increased the capitalthat the reclamation must defend and, on the other hand,they greatly decreased the time of concentration ofmeteorological waters (i.e. the time it takes rainfall watersin a certain point to reach the drains). The old drainagenetwork designed on the needs of agricultural landsresults insufficient. As a consequence, many inundationshave occurred causing great financial damages. Anexample is the major flood that affected the eastern partof the Padana Plain in the North of Italy during November

542 F. CAZORZI ET AL.

2010 (Marra et al., 2011), and that underlined how theagrarian landscape and the urbanization are connectedone to the other, and their deep connection represents acritical issue in flood risk management.In Northern Italy, the agrarian drainage network consists

generally of open main, sub-main, and lateral ditches,located both on the field boundaries but alsowithin the plots.Some areas lay below the mean sea level, and consequentlythe water fluxes on the network are either mechanically oralternately mechanically controlled by pumping stations.The whole network is subdivided into different drainagedistricts, each managed and monitored by a LandReclamation Consortium (LRC). Each LRC has theresponsibility to control, maintain and modernize thedrainage network, and to guarantee the hydraulic safety ofthe territory from flooding. This proper programming ofmeasures, especially for complex drainage systems, relieson a correct network characterization, starting from large-scale and up-to-date information about network positioningand geometry. Furthermore, drainage/reclamationservice criteria determine the requirements for hydraulicinfrastructure in the form of drainage density and storagecapacity. Storage capacity within the network, in particular,plays an important role in the design of drainage channelsand pumping stations, considering that a larger storage orretention capacity within the network system lowers therequirement for pumping capacity to achieve the sameservice objective (Malano and Hofwegen, 1999). Networkstorage capacity assessment, furthermore, is a crucial toolfor flood management: low values of channel storagecapacity can underline deficits in the network, and they canoutline areas whose hydrological behavior is potentiallycritical during floods.For the characterization of the network over large areas,

problems arise in Italy for two main reasons: i) there areno technical manuals with precise indications, valid forthe Country, regarding the suitable size and density of thein-field channel network, that remains a farmer choice;and ii) as in other European countries, each farm unit isvery often not continuous: as a result of land inheritanceand political action, properties are generally characterizedby a high degree of fragmentation, and each farmer’s landis divided into a collection of scattered plots. As aconsequence, when scaling up from an agricultural plotscale to the LRC scale, the structure of drainage networksappears to be highly variable in space and hard tocharacterize.For the localization of the network, if one does not

consider field surveys, the only available source is theofficial cartography (Regional Technical Maps – RTMs-at 1: 5000 scale). The provided information refers to thetime the map is produced. In Italy, the updating of theRTM is supposed to be done approximately every 10–15years, but no specific regulation exists: each regionalgovernment makes its own rules and specifications. It istherefore clear that users, when working with the RTM,must cope with the oldness of the derived information,that cannot account for all the frequent changes thedrainage network is subjected to.

Copyright © 2012 John Wiley & Sons, Ltd.

When dealing with the geometric characterization of thenetwork, channel geometries are generally homogeneousover each property, and they are related to the trenchers usedto build the channel, but no specific constraints exist thatrequire all the farmers to apply some specific machines.Consequently, channel geometries vary over the differentplots, and no information at all about channel widths orcross sections are available unless they are derived withlarge investments in field surveys (Lagacherie et al., 2004),with clear constraints for generalization.As a consequence, drainage length, drainage density and

storage capacity within the network are generally hard todefine, and tools for an accurate large-scale and up-to-datedetection of the agrarian network, and for an efficientcharacterization of its characteristics are deeply needed.Recent advances in data collection technology, such as

airborne and terrestrial laser scanning, enabled rapid,accurate, and effective acquisition of topographicinformation (Ackermann, 1999; Kraus and Pfeifer,2001; Slatton et al., 2007; Tarolli et al., 2009). Highresolution data are available concerning the groundelevation, and considering the artificial networkdimensions, i.e. width or height, this kind of data canbe a good candidate for agrarian landscapes’ linearelements mapping (Bailly et al., 2008). A first method todelineate artificial network reaches in agrarian landscapedirectly from the LiDAR point cloud has been assessedby Bailly et al. (2008). However, these authors underlinedhow their approach had two requirements: i) ditches areexclusively located at field boundaries, and ii) ageographic database of plot boundaries must be available.If both conditions are satisfied, ditch detection is definedfor some pre-located sites, and for those specificlocations, it is possible to estimate the ditches elevationfrom the scatter of LiDAR last pulse points. The twounderlined constraints make the method not feasible forlarge scale applications in the Italian context, where adense in-field network exists, the official cartographycannot be considered as a valid and up-to-date networkdatabase to identify specific sites, and field surveys forlarge areas are cost-prohibitive and time-consuming.How to objectively and automatically detect and

characterize the channel network in a floodplain context,where the network is highly variable and only fewinformation are available, is still an open question. Inother environmental contexts, automatic channel networkdetection benefits from the topographic information givenby high resolution (3m and better) LiDAR Digital TerrainModels (DTMs) (Tarolli and Dalla Fontana, 2009;Passalacqua et al., 2010; Pirotti and Tarolli, 2010; Sofiaet al., 2011). Nowadays LiDAR DTMs covering largeareas are readily available for public authorities, and thereis a greater and more widespread interest in theapplication of such information by agencies responsiblefor land management for the development of automatedmethods aimed at solving geomorphological andhydrological problems. In a floodplain context, problemsfor network extractions arise because the accuracy withwhich a DTM is able to detect and correctly represent the

Hydrol. Process. 27, 541–553 (2013)

543DRAINAGE NETWORK DETECTION IN AGRARIAN LANDSCAPE

hydrological asset of a catchment is determined by thestrength of the landscape gradient (i.e. flatness and/orslope changes). Low-relief areas, such as great alluvialplains, are therefore, more challenging even when high-resolution DTMs are available. For cultivated landscapes,drainage algorithms used on DTMs are unable torepresent anthropogenically modified overland flow paths(Duke et al., 2006; Garcia and Camarasa, 1999). Thedetection of the network in these contexts benefits frommorphological approaches, whereas channelized elementsare defined from local morphologies, and hydrologicalconnectivity is not accounted for.In this work, we propose a morphometric methodology

based on high resolution DTMs to (1) detect the drainagenetworks (ditches and channels) in agrarian/floodplaincontexts, and (2) to calculate some of the network summarystatistics (i.e. network length, width, drainage density andstorage capacity per unit of interest). The procedure startsfrom a grid DTM readily available for public authorities inItaly. The DTM (1m resolution) derives from a LiDARsurvey planned to have highly affordable costs: the surveycan be easily repeated during time, providing up-to-datelarge scale topographic information.The procedure is applied in two typical alluvial-plain

areas in the North East of Italy, and results are testedcomparing automatically derived network informationwith field surveyed ones.

STUDY SITES

For the present work, we identified two main study sites,selected as representative of the land reclamation drainagenetwork of the North East of Italy. The first site is a smallarea covering about 80 ha (a in Figure 1), where a detailedsurvey on channel cross sections has been done. To testthe large-scale effectiveness of the propose procedure, wealso applied the methodology to a wider site, coveringabout 30000 ha (b, Figure 1).

Site A: control area

The control area (Figure 1a) is a 82 ha site located in theFriuli Venezia Giulia region (North East of Italy), in themunicipality of Sacile (PN). The area contains a complexand dense artificial drainage network approximately 20 kmlong. The network (Figure 2) consists of open main, sub-main, and lateral ditches, located on the field boundaries butalso within the plots. The positioning of the ditches thatmainly compose this network is regular, and it closelydepends both on field subdivisions and agriculturalpractices. A detailed topographic survey was done, in orderto provide a broad database of typical cross sections,retrieving 170 measures of channel geometries (Figure 2).The survey was planned to obtain cross-section

measurement for any relevant change on the sections forevery significant network length (>30 m). A database ofbankfull widths (Wobs), mean depths and the cross-sectionalareas (Aobs) has been created. Ditches and channelsdimensions are characterized by widths ranging from 0.90to 6.10 m (average of 1.2 m for ditches and 2.4 m for


channels) and cross-sectional areas between 0.07 and 5.40m2 (average of 0.5 m2 for ditches and 0.8 m2 for channels).Being the drainage network artificial, its shape andcharacteristics are generally regular and homogeneouswithin the plots, and a strong relationship between the twoparameters is registered (Figure 3).Considering that the network is artificial and generally

built using trenchers with specific digging widths anddepths, it is possible to relate some specific cross-sectionalarea to network width, the assigned values in this case are0.4 m2 cross-sectional areas to widths lower than 2 m, 0.7m2 for widths up to 3 m and 1.5 m2 for sections larger than 3m. This characterization is area specific: it is an averagedcharacterization of cross-sections for widths ranges, and itmay vary in different areas according to the diggingtechniques adopted to create the ditches.

Site B: application area

The application area (Figure 1b) covers the tract of thelower Veneto plain (North-Eastern Italy) stretchingbetween Livenza and Tagliamento Rivers. The wholesurface (about 30000 ha) is regulated by Hydro-geologicalStructure Plan (PAI) and it belongs to the Lemene RiverBasin Authority whose borders outline the area easternboundaries. The west boundaries are underlined instead bythe Livenza river embankments.The application area is crossed from north to south by

several natural and artificial water collectors (Figure 1b).These watercourses originate from the northern part of thearea, in Friuli-Venezia Giulia, by waters emerging alongthe line of groundwater springs; they then flow throughVeneto receiving the superficial outflows. In the southernpart, the average ground elevation is at the sea level anddeclines to even 2 – 3 meters below it in the southernmostareas, close to the coast. For this reason, these water-courses in their final stretch become collectors of artificialoutflows, which originate from water scooping plants.Almost 60% of the area is artificially drained: at thepresent time, about 30 water scooping plants are working,with a total circulation of about 200 m3/s.The whole area includes different drainage units, based

mainly on a network of open channels and ditches, whichguarantees the drainage of meteoric water, and on acomplex embankments system, which prevent the wholearea from being flooded. For this study, we considered 50drainage units (Figure 1b), completely relying on artificialdrainage by water scooping plants. The drainage networkis mainly composed by minor channels and ditches, astypical for agricultural landscapes, and it spans for about7000 km, 30% of which offers just drainage functions.The remaining 70% is characterized by both irrigationaland drainage functionality making use of draining ditcheseither for irrigation or for drainage.

METHODOLOGY

The proposed morphological approach aims to a geometriccharacterization of the network (total length, drainage

Hydrol. Process. 27, 541–553 (2013)

Figure 1. Location of the control area (a) and the application area (b). Main drainage network as derived from official cartography is shown for both sites.For the application area (b), only channels and rivers are shown, and the considered drainage units (labelled from 0 to 49) are displayed


density, storage capacity within the channels for specificareas). The procedure does not consider a hydrologicalcharacterization: neither the connectivity of the network northe flow directions are accounted for.The method can be conceptualized in two main steps:

1. From a high resolution DTM, the network isautomatically detected and, always in an automaticfashion, a width and a length index are computed. Fromthese indexes, network length and network drainagedensity are computed directly. 2. In a second step, byassociating some characteristic cross-section areas tospecific width ranges, it is possible to derive the networkstorage capacity.While the first step is completely automatic, the second

step requires the user to identify roughly some widthranges and cross section shapes. Once these are defined,the network storage capacity is evaluated automatically.


The width and length indexes are evaluated directlyfrom a raster map, without any skeletonization procedure.This because even if algorithms for such operations areavailable in many contemporary GIS, they usuallyproduce highly irregular skeletons that are a poorrepresentation of the channel itself, and they requireadditional filtering and connection. However, consideringthat the DTM has a constraint due to the grid size (1 m)that is constant all over the map, two assumptions need tobe underlined: i) this constraint makes inapplicable thereading of cross section geometry directly on the DTM(Figure 4), but it allows an approximation of the ditcheswidths, considering that their minimum width is about 1m (equal to the DTM grid size); ii) an exact localcharacterization of the ditches is not expected, rather theaim is a characterization of the network for eachconsidered drainage unit.

Hydrol. Process. 27, 541–553 (2013)

Figure 2. Drainage network for the control area and the field surveyed cross-sections. Some examples of typical cross-sections are shown. Measuredcross-sectional areas are 0.46 m2 (Sect. n�1) , 1.97 m2 (Sect. n�6), 4.57 m2 (Sect. n�131) and 0.32 m2 (Sect. n�136)


From DTM to network detection and indexes evaluation

Drainage networks composed by ditches and channelsin agrarian landscapes are usually represented by linearand regularly distributed features. Land reclamationditches are commonly low ‘amplitude’ elements, con-sidering amplitude as the difference in relief between lowpoints on the ditch and height areas to either side, andthey are topographically subtle in comparison with theunderlying plain, especially for areas where the landscapegradient is relatively low. The approach we proposed isbased among consideration of having to extract localsmall-scale, low-relief features from the DTM and toeliminate as far as possible the large-scale landscape formsfrom the data. Similar approaches have been tested in otherenvironments for landform characterization (Carturan et al.,2009) and for feature extraction (Humme et al., 2006; Hiller


and Smith, 2008; Doneus and Briese, 2006). The core ideashared by these works is to apply a low-pass filter to theDTM, providing a smoothed elevation model representingan approximation of the large-scale landscape forms. Bysubtracting the DTM to this smoothed map, anapproximation of the local relief is achieved, whereonly small-scale topographic features are preserved, inspite of large-scale landscape forms. The derived model(Relative Elevation Attribute – REA- Carturan et al.,2009) represents a map of residual relief. This map isevaluated as

REAr ¼ �Er � EDTMr (1)

where �Er is the average elevation of cells within a circulararea with radius r (in meters) around the gridcell withelevation EDTM (Carturan et al., 2009).

Hydrol. Process. 27, 541–553 (2013)

Figure 4. Example of cross section geometry as depicted by the DTM (blue) compared to the surveyed ditch geometry (red)

Figure 3. Control area: surveyed cross-sectional areas (Aobs) and bankfull widths (Wobs) and linear relationship between the values


DTMs based on laser scanning are highly accurate,buildings as well as vegetation are normally alreadyfiltered. However, laser-based DTMs also show all man-made modifications of the earth’s surface such as largeroads, stamps of urban areas, high banks etc. Forextraction procedure based on relief control, prior tolocal relief evaluation, it is necessary to identify androughly mask these manmade terrain features on theoriginal DTM, at least where their density is high. Thismasking approach allows a better detection of the actualchannels, decreasing considerably false detections,avoiding peak values on the relief index due to artificialelements.Once the DTM has been roughly masked, Equation (1)

is applied, leaving only signature of the main topographicfeatures, detrending the landscape.In Figure 5 the DTM for a typical drainage unit is

shown (a). On the DTM some narrow linear elementrepresenting the drainage network are visible. In Figure 5bthe derived REA map is shown, highlighting the networkfeatures with higher values of the parameter. A profilegraph of the REA values for a surveyed section (5c) isprovided.To detect the network it is necessary to label the REA

peak values, and this is done through a thresholdingapproach. Tarolli and Dalla Fontana (2009) and Pirottiand Tarolli (2010) used a threshold identified as m-times


the standard deviation (s ) of curvature as an objectivemethod for channel network extraction from highresolution topography. In the proposed methodology,the chosen threshold is the standard deviation of REA(Figure 5b and 5c). The result (Figure 5d) represents aboolean map with features taking only binary values, 1 or0 for network pixels and landscape pixels respectively.This map is composed by clusters of pixels (groups of

connected cells having the same value). A single clusterdoes not necessarily imply the identification of a singlenetwork reach: each cluster might be a composed bymultiple reaches with different geometries. Figure 6shows an example of three different clusters: while clusterB and C correspond to single channels, each one with aspecific and constant width, cluster A comprises twoditches with different widths.Simple shape-related quantities, as area, width and

length can be consistently calculated for each cluster.Knowing the DTM grid size, the surface for each cluster(sc) simply refers to the sum of the areas of the pixelwithin the cluster.The network width index is computed for each cluster

according to the following steps.For each pixel on the cluster, the connected neighboring

cells along four directions D (0-180, 45-225, 90-270,135-315) are counted (Figure 6). The width wpx

associated to each pixel is

Hydrol. Process. 27, 541–553 (2013)

Figure 5. Drainage unit n� 23: input DTM (a), derived REA map and the threshold identified to detect the network (b), REA peak values correspondingto ditches (c). The boolean map detecting the drainage network derived after thresholding is shown (d)

Figure 6. Clusterization approach and cluster parameters evaluation. Cluster A is composed by reaches with different geometries, while cluster B and C aresingle reacheswith constant width. On the overview, an example of the pixel width evaluation is proposed: the red arrows show, for the considered pixel (reddot), the four directions D (0-180, 45-225, 90-270, 135-315) used for width evaluation. In this example, the derived pixel width for the pixel in analysis is 2

wpx ¼ min c0�180 ountpx; c45�225 ountpx; c90�270 ountpx; c135�315 ountpx� �

(2)


where Dcountpx is the number of pixels connected to theconsidered one (px) along the direction D. In theexample provided in Figure 6, for the pixel of interestthe evaluated wpx is 2.The width index Windex for each cluster is then

represented by the average value of pixel widths

Windex ¼

Pn

px¼1wpx

n(3)


where wpx derives from Equation (2) and n is the numberof pixel within the cluster.The length index Lindex is analytically derived as the

ratio between the cluster area and the cluster width index(Windex).

Lindex ¼ scWindex

(4)

Knowing both indexes for each cluster, it is possible toevaluate the areal drainage length (DLu) and drainage

Hydrol. Process. 27, 541–553 (2013)


density (DDu). Drainage Length is evaluated as the sumof Length indexes of the clusters belonging to theconsidered unit. Knowing the unit surface, drainagedensity (DDu) is evaluated as the ratio between drainagelength and the actual surface of the unit of interest.

Storage capacity within the network

Field surveying allows to define some representativecross-sectional areas for specific width ranges (see chap.2). This characterization is a user choice: it is an averagedcharacterization of cross-sections for widths ranges, and itmay vary in different areas according to the diggingtechniques adopted to create the ditches.This approach considers a simplified representations of

ditches cross-sections such as nested trapezoids, singletrapezoids, or rectangular cross-sections, as derived byfield survey. An optimal stream channel description shouldinclude detailed measurements of how depth varies withincreasing widths. However, due to the highly regularshape of agrarian drainage networks, this simplifiedestimation procedure is justified, and the averagingcharacterization accounts also for the differences thatmight be present over different plots due to farmer choices.It should be pointed out, anyway, that a) this process resultsin an average characterization of the network per unit ofinterest; and b) the relationship between width ranges andcross-sectional areas as it is formulated, cannot be appliedfor watersheds with a more natural network, such asmountainous watersheds of similar drainage area that arenot likely to have regular channel geometries.Once the width index is automatically derived, it is

possible to associate the user-defined cross-sectional areas(A) to the whole network, and the network storagecapacity for drainage unit (SCu) is defined as

SCu ¼

Pn

j¼1Lindexj �A� �

Su(5)

where Lindex is the length index for each cluster (j), Su isthe surface of the considered drainage unit, and A is theassociated cross sectional areas such as

A ¼ f Windex j

� �(6)

RESULTS AND DISCUSSION

Drainage networks as detected by the procedure and asmapped in the Regional Technical Map are shown for thewhole control area (a) and for a typical drainage unit forthe application area (b) in Figure 7.The accuracy of the methodology has been tested

considering some comparisons based on the availability ofactual network statistics. Whereas channel lengths andchannel geometries (bankfull width, depth and cross-sectional areas) are available in detail for the control area,acquiring channel geometry measurements for the whole


network of the application area (total length about 7000 km)would have been time-consuming and cost-prohibitive dueto the large extent of the site (30000 ha). Available referencenetwork statistics listed by sources are express in Table I.Considering the problems of the Regional Technical

Map, high-resolution digital aerial photos collected forthe application area during the same flight that providedthe LiDAR dataset, have been used in a time-consumingapproach to update manually the RTMs where needed,in order to provide a better information for qualityassessment.Assessment of drainage network positioning accuracy

was performed by calculating distances between thelocation of channels of the available reference data, andthe corresponding derived network. Median distance ofreference points to extracted drainage network for thestudy site is registered to be about 1 m (consistent withthe DTM grid size constraint). In some instances, theprocedure detects localized pixels, that do not correspondto the actual network location and that can be classified aseither noises or localized macro-depression on the soil.When compared to the actual land use, the proposedprocedure represents the network in its real conformation,while the original RTMs are not updated (Figure 8).A first quality assessment has been done considering

the width index (Windex) and the actual width derived fromfield survey (Wobs). One must note that, by definition, theproposed indexes are a unique averaged value for eachcluster of pixels, and each cluster does not necessarilyrepresent an individual reach, but it might be composedby multiple network reaches with different cross-sectiongeometries (see i.e. Figure 6, Chapt. 3.1). To test theeffectiveness of the procedure, some specific locationshave been considered, where the considered cluster wasactually a single reach of the network with a constantsurveyed width (Figure 9a). For the same clusters, thelength index (Lindex) has been compared to the fieldsurveyed length (Lobs) (Figure 9b).The result shows that when the considered cluster is

representative of a reach, the width index is generally also acorrect local estimation, with a measured RMS of about 0.2m. The same consideration can be done about the lengthindex: in this case, the RMS is estimated to be about 2 m.Using a transitive relationship, if the field measured

cross-sectional areas (Aobs) are strictly correlated to thefield measured widths (Wobs) (Figure 3), and the fieldmeasured width (Wobs) is in turn related to the width index(Windex) (Figure 9), then cross-sectional areas can be alsorelated to width index. By associating some characteristiccross-section area to specific width ranges, it is possible toautomatically derive the network storage capacity.When the storage capacity is compared to the field

surveyed one in the control area, a good concordanceamong measures is registered: the field surveyed storagecapacity for the area is about 156.14 m3/ha and the oneestimated by the procedure is 154.11 m3/ha. Theprocedure slightly underestimate the surveyed storagecapacity (SCu= 0.98 SCobs), and this is consistent withthe approximation in the width and length indexes

Hydrol. Process. 27, 541–553 (2013)

Table I. Reference statistics availability

Network reference statistics Control Area (80 ha) Application Area (30000 ha)

Network positioning DGPS field survey Manually Updated RTMsDrainage Length Field survey Manually Updated RTMsChannel Geometry Field survey (170 control points) Field survey (15 control point)Storage Capacity Field survey None

Figure 7. Reference drainage systems and network derived through the proposed approach for the control area (a) and for a typical drainage unit(drainage unit n� 23) for the application area (b)


(Windex = 0.98 Wobs and Lindex = 0.96 Lobs ). For theapplication area, storage capacity ranges between 50 m3/hato about 300 m3/ha: low values of channel storage capacitycan underline deficit in the network, and they can outlineareas whose hydrological behavior is potentially criticalduring floods.The regression technique was chosen to investigate the

relationship between evaluated network statistics (Drainagelength, drainage density and storage capacity) versus controlvalues considering at the same time values derived fromfield survey for the control area with the values derived fromthe updated Regional Technical Map for the applicationarea. The underlying assumption is that a similarityexists between the two sites. The main idea is to integratethe RTC-derived information with the field surveyed one,even if derived on a different site. The quality evaluation of


the methodology can be then estimated more accurately byusing the existing field surveyed values as an auxiliary base.Evaluated drainage lengths (DLu) have been tested

against drainage length derived from field survey (in thecontrol area, red dot in Figure 10) and updated fromcartography (application area, blue dots in Figure 10),highlighting how a strong linear relationship (R2 = 0.97)exists between the evaluated characteristic and themeasuredone in all cases. The evaluatedRMSE for the network lengthis estimated to be around 1.5 m and it is consistent with thevalues registered for the cluster length index.The values of evaluated drainage length and field

derived one for the control area behave in a similar waywhen considering length ranges between 0 and 60 km(Figure 10a). For the control area, the evaluatedDrainage Length is about 17 km, while the surveyed

Hydrol. Process. 27, 541–553 (2013)

Figure 8. Accuracy of evaluated network (green) and network as digitalized from official cartography (Regional Technical Maps –RTMs-) (red) beforeupdating, for drainage units n� 27 and 33. The underlying ortophoto shows how official cartography needs some update and adjustment

Figure 9. Comparison between a) width (Windex ) and b) length (Lindex) indexes and the correspondent surveyed measures (Wobs and Lobs)


one is 19.50 km. In this case the procedure slightlyunderestimate the measures.When evaluating drainage density (Figure 11),

correlation between the measured values and the evaluatedones is less strong, but still valuable (R2 = 0.67). Testing R2

with two-tailed Student’s t-tests for statistical significance


and a=0.05, anyway, shows that this value represents asignificant linear relationship between the two variables forthe considered datasets.As mentioned before, due to the large extent of the

application area, it was not possible to have detailedinformation about storage capacity within the network,

Hydrol. Process. 27, 541–553 (2013)

Figure 10. Drainage length derived by the proposed procedure (DLu) and field surveyed Drainage Length (DLobs) for the control area (red dot) and foreach one of the drainage units of the application area (blue dots)

Figure 11. Drainage density derived by the proposed procedure (DDu) and field surveyed Drainage density (DDobs) for the control area (red dot) and foreach one of the drainage units of the application area (blue dots)

Figure 12. Relationship between the Storage Capacity (SCu) derived by the proposed procedure and the evaluated Drainage Density (DDu) for thecontrol area (red dot) and for each one of the drainage units of the application area (blue dots)


but for this study case, regression techniques, anyway,indicated that the evaluated storage capacity is stronglyrelated (R2 = 0.83) to drainage density (Figure 12), thathas been proven to be related to the correspondent controlvalues (Figure 11).The strength of correlation between REA derived

values and the surveyed ones, in the majority of theinstances, suggests a high reliability of the method.


CONCLUSIONS

We presented a method for fast drainage networkdetection and quantification of water storage capacityand other network main statistics for agrarian landscapewithin floodplains. In this context, flow-direction matricesare usually not applicable to extract the network, due tothe low gradient of the surface. At the same time, for such

Hydrol. Process. 27, 541–553 (2013)


vast areas, gathering detailed information about thenetwork can be time-consuming and cost-prohibitive.The performance testing has demonstrated that thisimplementation can be efficient on gathering detailedand correct information for large datasets, thanks to themodern mapping technologies that can rapidly producelarge DTMs at high resolutions. The accuracy assessmentusing field surveyed control values and informationderived from a manually updated cartography providedinsight into the accuracy of drainage network derivedthrough the proposed procedure. The results suggest thatthe conceptually simple approach to remove largelandscape features to extract small linear elements issuitable, fast and enough accurate for this kind ofenvironment. The strength of correlations betweenderived values and the control values in the majority ofthe instances suggests a promising and valuable reliabilityof the method, especially considering the regularity ofshapes and structures of agrarian drainage network.Some small inconsistencies between values exist, due

to the DTM grid cell constraint that determines a slightunderestimation of the derived parameters. However, thenetwork extracted through the methodology has beenproven to be more reliable than the information providedby the cartography, when updates on the maps are notavailable. The method, furthermore, allows the definitionof indexes generally available only through time-consuming and cost-prohibitive field surveys (networkgeometry): it was possible to identify correctly somenetwork characteristics over a large extent (30000 ha)and, using only few control points (15), to deriveinformation previously unknown about the storagecapacity within the channels. This assessment is a crucialtool for flood management: low values of channel storagecapacity can underline deficit in the network, and they canoutline areas whose hydrological behavior is potentiallycritical during floods.This fast processing of large DTM datasets opens their

application to a new level of detail and spatial extent, forexample in rapid response operations and they can becomea useful tool in regions where available cartography needsto be updated or correct according to the land-use changes,especially in agrarian/urban contexts.

ACKNOWLEDGEMENTS

Analysis resources were provided by the InterdepartmentalResearch Center for Cartography, Photogrammetry,Remote Sensing and GIS, at the University of Padova—CIRGEO. LiDAR data for the Control Area were providedby the Municipality of Sacile (PN).

REFERENCES

Ackermann F. 1999. Airborne laser scanning – present status and futureexpectations. ISPRN Journal of Photogrammetry and Remote Sensing54: 64–67.

Bailly JS, Lagacherie P, Millier C, Puech C, Kosuth P. 2008. Agrarianlandscapes linear features detection from LiDAR: application to


artificial drainage networks. International Journal of Remote Sensing29(12): 3489.

Brath A, Montanari A, Moretti G. 2003. Assessing the effects on flood riskof land-use changes in the last five decades: an Italian case study. IAHSPublication, 278. IAHS Press: Olanda; 435–441.

Carluer N, Marsily GD. 2004. Assessment and modelling of the influenceof man-made networks on the hydrology of a small watershed:implications for fast flow components, water quality and landscapemanagement. Journal of Hydrology 285: 76–95.

Carturan L, Cazorzi F, Dalla Fontana G. 2009. Enhanced estimation ofglacier mass balance in unsampled areas by means of topographic data.Annals of Glaciology 50: 37–56.

Dagès C, Voltz M, Bsaibes A, Prévot L, Huttel O, Louchart X, Garnier F,Négro S. 2009. Estimating the role of a ditch network in groundwaterrecharge in a Mediterranean catchment using a water balance approach.Journal of Hydrology 375: 498–512.

Doneus M, Briese C. 2006. Full-waveform airborne laser scanning as atool for archaeological reconnaissance. In From Space to Place. 2ndInternational Conference on Remote Sensing in Archaeology, CampanaS, Forte M (eds). BAR International Series 1568. Archaeopress:Oxford; 99–105.

Duke G, Kienzle S, Johnson D, Byrne J. 2006. Incorporating ancillarydata to refine anthropogenically modified overland flow paths.Hydrological Processes 20: 1827–1843.

Dunn SM, Mackay R. 1996. Modelling the hydrological impacts of openditch drainage. Journal of Hydrology 179: 37–66.

Gallart F, Llorens P, Latron J. 1994. Studying the role of old agriculturalterraces on runoff generation in a small Mediterranean mountainousbasin. Journal of Hydrology 159: 291–303.

Garcia M, Camarasa A. 1999. Use of geomorphological units to improvedrainage network extraction from a dem: Comparison betweenautomated extraction and photointerpretation methods in the carraixetcatchment (valencia, spain). International Journal of Applied EarthObservation and Geoinformation 1(3-4): 187–195.

Gascuel-Odoux C, Aurousseau P, Doray T, Squividant H, Macary F,Uny D, Grimaldi C. 2011. Incorporating landscape features to obtain anobject-oriented landscape drainage network representing the connect-ivity of surface flow pathways over rural catchments. HydrologicalProcesses 25: 3625–3636.

Hiller JK, Smith M. 2008. Residual relief separation: digitalelevation model enhancement for geomorphological mapping.Earth Surface Processes and Landforms 33: 2266–2276. DOI:10.1002/esp.1659.

Humme A, Lindenbergh R, Sueur C. 2006. Revealing Celtic fields fromlidar data using kriging based filtering. Proceedings of the ISPRSCommission V Symposium, Dresden, 25–27 September, Vol. XXXVI,part 5.

Kraus K, Pfeifer N. 2001. Advanced DTM generation from LIDAR data.International Archives of Photogrammetry and Remote SensingXXXIV-3/W4: 23–35.

Lagacherie P, Diot O, Domange N, Gouy V, Floure C, Kao C, Moussa R,Robbez-Masson J, Szleper V. 2004. An indicator approach fordescribing the spatial variability of human-made stream network inregard with herbicide pollution in cultivated watersheds. EcologicalIndicators 6: 265–279.

Levavasseur F, Bailly JS, Lagacherie P, Colin F, Rabotin M. 2011.Simulating the effects of spatial configurations of agricultural ditchdrainage networks on surface runoff from agricultural catchments.Hydrological Processes, Accepted Article. DOI: 10.1002/hyp.8422

Malano H, Hofwegen PV. 1999. Management of Irrigationand DrainageSystems: A Service Approach. Monography IHE 3. A.A. Balkema:Rotterdam, Netherlands. ISBN 90 5410 482 1.

Marofi S. 1999. Rôle des échanges nappes-fossés dans le fonctionnementhydrologique d’un bassin versant en milieu méditerranéen cultivé. PhDthesis, Ecole Nationale Supérieure Agronomique de Montpellier,France; 240.

Marra F, Zanon F, Penna D,Mantese N, Zoccatelli D, Cavalli M,MarchiM.2011. Hydrometeorological and hydrological analysis for the Nov 1,2010 flood event in North-eastern Italy. Geophysical Research Abstracts13: EGU2011-10727-1.

Moussa R, Voltz M, Andrieux P. 2002. Effects of the spatialorganization of agricultural management on the hydrological behaviourof a farmed catchment during flood events. Hydrological Processes16(2): 393–412.

Passalacqua P, Tarolli P, Foufoula-Georgiou E. 2010. Testing space-scalemethodologies for automatic geomorphic feature extraction from lidarin a complex mountainous landscape. Water Resources Research 46:W11535. DOI: 10.1029/2009WR008812.

Hydrol. Process. 27, 541–553 (2013)


Pirotti F, Tarolli P. 2010. Suitability of LiDAR point density and derivedlandform curvature maps for channel network extraction. HydrologicalProcesses 1187–1197. DOI: 10.1002/hyp.7582.

Slatton KC, Carter WE, Shrestha RL, Dietrich WE. 2007. Airborne laserswath mapping: achieving the resolution and accuracy required forgeosurficial research. Geophysical Research Letters 34: L23S10.

Sofia G, Tarolli P, Cazorzi F, Dalla Fontana G. 2011. An objective approachfor feature extraction: distribution analysis and statistical descriptors forscale choice and channel network identification. Hydrology and EarthSystem Sciences 15: 1387–1402. ISSN: 1027-5606.DOI: 10.5194/hess-15-1387-2011.


Tarolli P, Arrowsmith JR, Vivoni ER. 2009. Understanding earth surfaceprocesses from remotely sensed digital terrain models. Geomorphology113: 1–3.

Tarolli P, Dalla Fontana G. 2009. Hillslope to valley transitionmorphology: new opportunities from high resolution DTMs.Geomorphology 113: 47–56.

Voltz M, Andrieux P, Bouzigues R, Lennartz B, Louchart X,Moussa R, Ribolzi O. 1998. Transport of water and matter in afarmed sedimentary soil system in the mediterranean environment.Proceedings of the 16th International Congress of Sol Science,Montpellier; 8.

Hydrol. Process. 27, 541–553 (2013)

Documents

Drainage Network Detection and Assessment of Network Storage Capacity in Agrarian Landscape