You are on page 1of 11

WATER RESOURCES RESEARCH, VOL. 33, NO.

2, PAGES 309-319, FEBRUARY 1997

A new method for the determination of flow directions

and upslope areas in grid digital elevation models


David G. Tarboron
Utah Water ResearchLaboratory,Utah State University,Logan

Abstract. A new procedurefor the representationof flow directionsand calculationof


upslopeareasusingrectangulargrid digital elevationmodelsis presented.The procedure
is basedon representingflow direction as a singleangle taken as the steepestdownward
slopeon the eight triangularfacetscenteredat each grid point. Upslope area is then
calculatedby proportioningflow betweentwo downslopepixelsaccordingto how closethis
flow directionis to the direct angleto the downslopepixel. This procedureoffers
improvementsover prior proceduresthat have restrictedflow to eight possibledirections
(introducinggrid bias) or proportionedflow accordingto slope(introducingunrealistic
dispersion).The new procedureis more robustthan prior proceduresbasedon fitting
local planeswhile retaining a simplegrid basedstructure.Detailed algorithmsare
presentedand resultsare demonstratedthroughtest examplesand applicationto digital
elevation data sets.

Introduction of grid digital elevation models,pointing out their strengths


and weaknessesand givingthe motivationfor a new method. I
Flow directionsbasedon digital elevationmodels(DEMs) then describethe new procedurefor calculationof flow direc-
are needed in hydrology to determine the paths of water, tionsand the procedurefor calculationof upslopearea on the
sediment,and contaminantmovement.Two importantdistrib- basisof the new flow directions.Illustrativeexamplesare used
uted quantitiesthat dependon flow directionsare the upslope to compare the new method with existing methods. These
area and specificcatchmentarea. Upslope area,A, is defined includesimpletest cases,where the theoreticalresultis known
as the total catchmentarea above a point or short length of and bias and squareerror can be evaluated,aswell aslow- and
contour[Mooreet al., 1991].The specificcatchmentarea,a, is high-resolutionDEM data where the evaluationis qualitative.
definedasthe upslopearea per unit width of contour,L, (a = Readersfamiliar with the issuessurroundingthe computation
A/L) [Mooreet al., 1991]and is a distributedquantitythat has of upslope area based on grid DEMs could skip the back-
importanthydrological,geomorphological, and geologicalsig- ground section.
nificance[Costa-Cabraland Burges,1994]. The specificcatch-
ment area contributingto flow at any particular location is
useful for determining relative saturation and generation of
Background
runoff from saturation excess in models such as TOPMODEL Grid DENs consist of a matrix data structure with the to-

[Bevenand Kirkby,1979;Bevenet al., 1984;Woodet al., 1990]. pographicelevationof eachpixelstoredin a matrixnode.Grid


Specificcatchmentarea togetherwith other topographicpa- DENs are distinctfrom other DEM representationssuchas
rametershasalsobeen usedin the analysisof processes suchas triangular irregular network (TIN) and contour-baseddata
erosionand landslides[Dietrichet al., 1992, 1993; Wu, 1993; storagestructures.Grid DENs are readilyavailableand simple
Montgomeryand Dietrich, 1994]. Upslope area is commonly to use and hence have seen widespreadapplicationto the
usedfor the automaticdemarcationof channelsrelyingon the analysisof hydrologicproblems[Mooreet at., 1991].However,
notion of a criticalsupportarea [O'Callaghanand Mark, 1984; they sufferfrom somedrawbacksthat arise from their grided
nature.
Jensonand Domingue,1988;Morris and Heerdegen,1988; Tar-
boton, 1989; Lammers and Band, 1990; Tarboton et al., 1991, The earliestand simplestmethod for specifyingflow direc-
1992;Martz and Garbrecht,1992]. Judgingby the number of tions is to assignflow from each pixel to one of its eight
recentpapers,there is considerablehydrologicinterestin the neighbors, either adjacent or diagonal, in the direction with
effectof grid scaleand proceduresfor computationof specific steepestdownwardslope.This method, designatedD8 (eight
catchmentarea [Zhang and Montgomery,1994; Quinn et al., flow directions),was introducedby O'Catlaghanand Mark
1995;WolockandMcCabe,1995].It is thereforeimportantthat [1984] and has been widely used [Marks et at., 1984; Band,
flow directionsand specificcatchmentareasbe accuratelyde- 1986; Jensonand Domingue, 1988; Mark, 1988; Morris and
termined free from grid artifacts. Heerdegen,1988; Tarboronet at., 1988; Tarboron,1989;Jenson,
1991;Martz and Garbrecht,1992]. In the contextof a grid the
In this paper a new method for calculatingflow directions,
upslopearea,A, is the area contributingto eachpixel and may
upslope,and specificcatchmentareas is presented.In what
be estimatedas the product of the number of pixelsdraining
follows I first review the currently availableproceduresfor
througheachpixel and pixel area. The specificcatchmentarea,
calculatingupslopeand specificcatchmentareason the basis
a, is then estimatedasA/L, takingL as the pixel width.
Copyright 1997bytheAmerican GeophysicdlUnion. The D8 approachhas disadvantages arisingfrom the dis-
Paper number 96WR03137. cretizationof flow into only one of eight possibledirections,
0043-1397/97/96WR-03137509.00 separatedby 45 [e.g.,Fairfieldand Leymarie,1991; Quinn et
309
310 TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS

a) North North
13 9
:
5
J
9 11 b)13 9 5 9 11
.........
-0:$.......................
a-......................
-si.................
-'e-1-75
.........
.......
':$
....................
*.........
"T.....
-sis.
.................
-i-$
.........
11 i i'"1-1 i 12 I
West----O:7$
................
-77H
.....................
1:
....................
-1---3
........
East Wesr"'8:7$
..............
-7:.'.-75
.................
-t0....................
-1--,1.---
I
......'East
C F
I

.........
7:'3'75
................
0d'25 .................
t"3'
:95...............
't7-J25
........
........
7:-375
...........
-"/'-6
:i-25
......
.xu:'..
.....
-1-'3'i95
................
-t-7-i-25
.........
8 6 " 5.8 30 12 8 6 5.8 30 12
South South

Figure 1. HypotheticalDEM subseton a block-centeredgrid. Corner elevationsare calculatedby averaging


the elevationsof adjoiningcells.(a) Flow directionsdeterminedusinglocallyfitted planes[Lea, 1992]. (b)
Flow directionsby thispapersnew procedure.In Figure la considerflow from the pixelwith centermarked
A. The pixelimmediately
to thenorthisthelowestneighbor,
withelevation5, sothatiswherewaterfromthis
pixel shouldflow. Somewater from pixel A shouldalso flow to the southor southwest, sincepixel A is a
saddle.Lea'smethodhaswater flow to the northwestalongthe directionof the planeslopeto point E, where
it crossesinto pixel B. The edgebetweenpixelsA and B is a discontinuous edgewith flow from the adjacent
pixelsin opposingdirections,a situationthat posesproblemsfor planeflow algorithms.Flow from A is routed
southalongthe edgefrom E to F, where it crossesinto pixel C and without exceptioncheckinggetsstuckin
an infiniteloop,movingfrom pixel C to D to A to B aroundthe point F. Furthermorethe flow originatingat
the centerof pixel B goessoutheastto point H, whereit enterspixelA and then flowsnorth alongthe edge
(sincethe slopevectorcenteredat A hasa north component),to point I, in an uphill direction,whereit enters
the pixelJ. Thusflowsfrom B andA movein oppositedirectionsalongthe sameedge,an unrealisticsituation.
Another problemis evidentat pixel K. Here flow hasa large north componentdue to the elevation30 to the
south,despitethe fact that its lowestneighboris to the southeast,and neighborwith steepestslopeis to the
east.

al., 1991; Costa-Cabraland Burges,1994]. Fairfield and Ley- tinuously(as an anglebetween0 and 2') and withoutdisper-
marie [1991] suggestedovercomingthese problemsby ran- sion. Costa-Cabraland Burges[1994] presentedan elaborate
domly assigninga flow direction to one of the downslope set of procedures,named DEMON (digital elevationmodel
neighbors,with the probabilityproportionalto slope. networks),that extendsthe ideasof Lea [1992].Grid elevation
Multiple flow direction methods[Quinn et al., 1991;Free- values are used as pixel corners,rather than block centered,
man, 1991]havealsobeen suggested asan attemptto solvethe and a plane surfaceis fitted for eachpixel. Costa-Cabraland
limitationsof D8. Theseallocateflow&actionallyto eachlower Burges[1994] recognizedflow as two-dimensionaloriginating
neighborin proportionto the slope (or in the caseof Free- uniformlyover the pixel area, rather than trackingflow paths
man's method, slope to an exponent)toward that neighbor. from the centerpoint of eachpixel.They evaluateupslopearea
Multiple flow directionmethods,designatedhere by MS (mul- throughthe constructionof detailedflow tubes.
tiple directionsbasedon slope), have the disadvantagethat The assumptionof a plane fit locallyto eachpixel requires
flow from a pixel is dispersedto all neighboringpixelswith approximationbecauseonly three points are required to de-
lower elevation. termine a plane. The best fit plane cannot in generalpass
Dispersionis inherentin any method (includingthe one I throughthe four corner elevations,leadingto a discontinuous
describebelow) that assigns flow from one pixel to more than representationof the surfaceat pixeledges.Planesfit locallyto
one downslopeneighbor and manifests itself in terms of certain elevation combinations can lead to inconsistent or
spreadingof flow from a singlepixel. It couldbe arguedthat counterintuitiveflow directionsthat are a problem in both
this doesnot matter becausethe modelsthat usea may use it Lea's[1992]methodandin DEMON. Figure1 illustratessome
as a surrogatefor a physicalquantitythat is affectedby disper- of theseproblems.Theseproblemsillustratedin the contextof
sion.However,dispersionis inconsistent with the physicaldef- Lea's method are also present in DEMON, sincethe same
initionof upslopearea,A, andspecificcatchmentarea,a. It is plane flow directionswould arise giventhe corner elevations
important,to the extentpossible,to minimizedispersionin the shownin Figure 1. The codingof approachesbasedon fitting
calculationof a. Then, if necessary, physicaldispersioncanbe localplanes,suchas Lea's methodand DEMON, so that they
modeledseparately. are robust and work for all possibleelevationcombinations
Lea [1992] developedan algorithmthat usesthe aspectas- that may arisein real data is difficult.There are many excep-
sociatedwith each pixel to specifyflow directions.Flow is tions, suchas the one illustratedin Figure 1, that need to be
routed asthoughit were a ball rolling on a plane releasedfrom anticipatedand specialcodedevelopedto accountfor them. In
the center of each grid cell. A plane is fit to the elevationsof fact, the codefor DEMON upslope[Costa-Cabral andBurges,
pixel corners,thesecornerelevationsbeingestimatedby aver- 1994]is unavailablebecauseit is "hard to programand full of
agingthe elevationsof adjoiningpixel centerelevations.This special cases"(M. Costa-Cabral, personal communication,
procedurehasthe advantageof specifyingflow directioncon- 1995).
TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS 311

Column indices i
i i

j-1 j j+l i i

i i
Proportionof flow to pixel. - -, i i

(i-l,
'
j)isOt2/(Otl,-I-Ot2).
Proportion
of
flowtopixel
,
i
i
i

,, . ,, m, ' (i-1 j+l) is i

i ' '3' "


'
Flowdirection
measured as 0
counter-clockwise
, anglefrom east.
!

Steepest
downslope
dire,ction

w
Facet
numbering
i
i i i

i i i

i i i Figure 3. Definition of variablesfor the calculationof slope


on a singlefacet.
Figure 2. Flow directiondefinedassteepestdownwardslope
on planartriangularfacetson a block-centeredgrid.
2r) is determinedin the directionof the steepestdownwards
slopeon the eight triangularfacetsformed in a 3 x 3 pixel
The motivationfor developinga newmethodstemmedfrom
windowcenteredon the pixel of interest.The useof triangular
frustrationwith the way theseexistingproceduresworked.The
facetsavoidsthe approximationinvolvedin fitting a plane and
followingissuesare relevant to the evaluationand designof
the influenceof higher neighborson downslopeflow. Where
DEM flow direction procedures:(1) the need to avoid or
the directiondoesnot followone of the cardinal(0, r/2, r, and
minimizedispersion;(2) the needto avoidgrid bias,due to the
3r/2) or diagonal(r/4, 3r/4, 5r/4, and 7r/4) directions,up-
orientationof the numericalgrid; (3) the precisionwith which
slopearea is calculatedby apportioningthe flow from a pixel
flow directionsare resolved;(4) a simple and etficientgrid
betweenthe two downslopepixelsaccordingto how closethe
basedmatrix storagestructure;and (5) robustness, the ability
flow angleis to the directangleto that pixelcenter.Sinceonly
to copewith "difficult"data,suchasthe saddlein Figure 1 but
a singlenumberneed be savedfor eachpixel to representthe
alsowith pits and flat areas.
flow field, computerstorageis simpleand efficient.Somedis-
The D8 methoddoeswell on points1, 4, and 5 but resolves
persionis introducedby the proportioningof flow between
flow directionstoo coarsely(point 3), introducinggrid bias
downslopepixels,but this is minimized since flow is never
(point 2). The randomnessin the method of Fairfield and
proportionedbetweenmore than two downslopepixels.
Leymarie[1991]is not appealing.Upslopeand specificcatch- Figure 2 illustrates the calculation of flow directions.A
ment areasare deterministicquantitiesthat we shouldbe able
block-centeredrepresentationis used with each elevation
to computein a repeatableway.The MS methodavoidsgrid
value taken to representthe elevationof the center of the
bias(point2) but introducessubstantial dispersion.Sinceflow
corresponding pixel.Eight planartriangularfacetsare formed
may be proportionedin up to eightpossibledirections,eight
betweenthe pixel and its eightneighbors.Each of thesehasa
numbersneedto be stored(or recalculatedeachtime they are
downslopevectorwhichwhen drawnoutwardfrom the center
needed) for each pixel, resultingin inefficientdata storage
may be at an angle that lies within or outside the 45 (,r/4
(point 4). The plane flow methods[Lea, 1992; Costa-Cabral
radian)anglerangeof the facetat the centerpoint.If the slope
and Burges,1994] are appealingin that they are deterministic
vectorangleiswithinthe facetangle,it represents the steepest
and preciselyresolveflow directions.However,they are sus-
flow directionon that facet.If the slopevectorangleis outside
ceptibleto problems(point 5) arisingfrom the approximation
a facet,the steepestflow directionassociated with that facetis
involvedin fitting a plane throughfour points.
taken along the steepestedge.The flow directionassociated
with the pixel is taken as the directionof the steepestdown-
Calculation of Flow Directions slopevectorfrom all eight facets.
To implementthisprocedurefirst considera singletriangu-
Here I proposea new methodfor the representationand
calculationof flow directionsthat is a compromiseon the lar facet (Figure 3). Slope (downward)is representedby the
issuesraisedin the previoussection.It recognizesthe advan-
vector(Sl, s2) where
tagesof Lea's [1992]methodandDEMON throughthe assign- Sl-- (eo- el)/dl (1)
ment of a singleflow directionto each cell. This singleflow
direction(representedasa continuousquantitybetween0 and s2= (el- e2)/d2 (2)

Table 1. Facet Elevation and Factorsfor Slope and Angle Calculations


Facet 1 2 3 4 5 6 7 8

eo eid eid ed eid eid eid eid eij


el eij+ 1 ei- 1d ei- ld eij- 1 eid- 1 ei+ ld ei+ 1d eld+ 1
e2 el- ld+ 1 el- 1,j+ 1 el- 1,j- 1 el- 1,j- 1 el+ ld- 1 el+ 1,j- 1 el+ 1,j+ 1 el+ 1,j+ 1
ac 0 1 1 2 2 3 3 4
az 1 -1 1 -1 1 -1 1 -1
312 TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS

A. Theoretical. B. D8.

"....... .?'.' :: i::;


........

.... :........:.
..........................
. ......................
. ....
....
. ............ ......
, ,,
........... ::::::::::::::::::::::
.....

'" ';- 3' .... '-2{


'"?'":":":'
'":"':

C. MS. D. Lea's method

E. DEMON F. Doo

Figure 4. Top half of an outwarddrainingcircularcone.Elevationwasdefinedas200 minusthe radiusfrom


the center on a 16 x 16 grid with grid spacing10 units. Specificcatchmentarea is theoreticallyradius/20
rangingfrom 0 at the centerto 53 on the corners.Contours(10-unit interval) depictelevation.Gray scale
depictscontributingarea(1 whiteto 60 black).(a) Theoreticalvalues,(b) singledirection(D8) procedure,(c)
Quinnet al.'s[1991]procedure(MS), (d) Lea's [1992]method,(e) DEMON [Costa-Cabral andBurges,1994],
and (f) new procedure(D).

and ei and d are elevationsand distancesbetweenpixelsas ra = air' + acVr/2 (6)


labeled in Figure 3. The slopedirectionand magnitudeare
r = tan- (s2/s0 Themultiplieraf andconstant
ac dependonthefacetselected
(3) and are listed in Table 1. The procedurethat searchesfor the
S = (S12
-'1
c S22)
1/2 facetwith the largestslopeproceedsin the order of facets1 to
If r isnotin therange(0, tan- (d2/d)), thenr needsto be 8 shownin Figure 1 and in the caseof ties (facetswith equal
set as the directionalongthe appropriateedgeand s assigned slope)picksthe first one. In natureties are extremelyrare so
as the slopealong that edge. the bias introducedby this is deemednegligible.
In the casewhere no slopevectorsare positive(downslope)
if r < 0, r = 0, s = s (4) a flow directionangleof -1 is usedto flag the pixel as "unre-
if r > tan- (d2/di), r = tan- (d2/di), solved,"that is, a flat area or pit. Unresolvedflow directions
(5) are resolvediterativelyby makingthem flow towarda neighbor
s = (eo- e2)/(d
2 + d22)
1/2 of equalelevationthat hasa flow directionresolved.This is the
sameapproachfor resolvingpits and flats as used in the D8
Next recognizethat eachof the eightfacetsdepictedin Figure
method[e.g.,Mark, 1988;JensonandDomingue,1988].I there-
2 canbe mappedby appropriateselectionof cornerelevations
and rotation/transformationonto the facet in Figure 3. Table 1 fore usethe calculationof D8 flow directionsasa preprocessor
givesthe nodeelevationscorresponding to the cornersof each to raise the elevationof all pixelsin a pit to the level of the
of the triangular facetsused to calculateslopesand anglesin overflow.Then where pixelsare flaggedas "unresolved,"the
(1)-(5). Theseare arrangedsuchthat eo is the centerpoint,e flow anglereturnedby the D8 procedureis used.This ensures
is the point to the side,and e2 is the diagonalpoint.The local that flat pixelsdrain to a neighborthat ultimatelydrainsto a
angle associatedwith the largestdownwardsslope from the lower elevation,eliminatingthe possibilityof inconsistencies
eight facets (r' = r from facet with maximums) is then suchasloopsin the flow directionangles.This methodof repre-
adjustedto reflectan anglecounterclockwise from east(Figure sentingflow directionsbasedon triangularfacetsis designated
2) to obtainthe flow directionangle. Dc (an infinitenumberof possiblesingleflow directions).
TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS 313

A. Singledirectionprocedure,
D8 B Quinnet al. (1991)procedure,
MS C. Lea's(1992)method

D. DEMON E. New Procedure, Doo


.........

Figure 5. Circular cone influencemaps for the circledpixels.Gray scalerangesfrom white (0 or no


influence)to black(1 or 100% influence).

Calculation of Upslope Areas The calculation is initiated by calling this function for the
Upslopearea is calculatedusinga rccursivcprocedurethat outlet pixel. It then recursivelycalls itself for all pixels that
is an extensionof the very efficient rccursivcalgorithm for contribute to the upslope area at the outlet. The recursion
singledirections[Mark, 1988].The upslopearea of eachpixel stopswhen it reachesa pixel that has no pixelsupslope.
is taken as its own area (one) plusthe area of upslopeneigh-
borsthat havesomefractiondrainingto the pixel in question. Illustrative Examples
The flowfrom eachcelleitherall drainsto oneneighbor(if the This sectiongivesexamplesof resultsfrom this method,D%
anglefalls alonga cardinalor diagonaldirection)or is on an comparedto the singledirectionapproach,D8; Quinn et al.'s
anglefalling betweenthe direct angleto two adjacentneigh- [1991]multidirectionalgorithm,MS; Lea's [1992]method;and
bors.In the latter casethe flow is proportionedbetweenthese DEMON [Costa-Cabral and Burges,1994].In theseexamples
two neighborpixelsaccordingto how closethe flow direction we use the notion of influenceand dependencemaps. The
angle is to the direct angle to those pixels. The following influencefunctionI(x, Xo) is definedas the upslopearea at
pseudocodegivesthe logic of this algorithm: eachpixelx from a specificpixelXo. It mapswhere flow from
ProcedureDPAREA(i, j) pixelxo goesand how it is dispersed.It is computedby running
if AREA(i, j) is known a modified versionof the procedurefor calculatingupslope
then area that usesan area contributionof one from pixelx o but
no action zero for all other pixels.The dependencefunctionD(x, Xo) is
else the oppositeof the influencefunction,definedasD(x, Xo) =
AREA(/, j) = 1 (the area of a singlepixel) I(xo, x). The upslopearea at pixelxo is composedof the sum
for eachneighbor(locationin, jn) of the area of upslopepixelsthat havesomeproportionof their
p = proportionof neighbor(in, jn) that drainsto flowgo throughpixelxo.D(x, Xo) mapsthe contributionfrom
pixel (i, j) basedon angle pixelx to the calculationof upslopearea at Xo. It is calculated
if (p > 0) then throughrepeated evaluationof the influencefunction.
call DPAREA(in, jn) (this is the recursivecall to Figure 4 showsthe upslope area by each approachfor a
calculatearea for the neighbor) circularcone.Figure 5 showsthe influencemapsfrom eachof
AREA(i, j) = AREA(i, j) + p x the five algorithms(D8, MS, Lea's [1992] method,DEMON,
AREA(in, jn) and Din) appliedto the circularcone.D8 resultsin no spread-
Return ing, but flow paths(whichare what the influencemap plots)
314 TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS

are constrainedto grid directions.MS followsthe topographic


slope but introduces substantialdispersion.Lea's [1992]
methodmarches
downthecontours
in stairstepfashion,while
DEMON andD strikea balancewith spreadingslightlywider
thandivergence perpendicularto the contours.For somepixels
(e.g.,the lowerrightonein Figure5e) D resultsin no spread-
ing. This is a pixelwhere the flow directionis alignedwith the
grid diagonal. In general, when the topographicslope is

A. Singledirectionprocedure,
D8.----.""-
lB. MS

B.Quinn__.et
al.(1991)
procedure,
MS
IC. Lea's

-C. Lea's (1992)

D.
DEMO
_NZ

E. New Procedure,Doo.'---

Figure 6. Inward cone. Contours show elevation. The left


panelsshowupslopearea in grayscale.The right panelsshow
the influencefunctionfrom the circledpixel. (a) D8 method,
(b) MS method[Quinnet al., 1991],(c) Lea's [1992]method,
(d) DEMON [Costa-Cabraland Burges,1994], and (e) D
method. Figure 7. Dependencemapsfor planarsurface.Gray scale(0
white, 1 black) indicatesthe fractionof eachpixel upslopeto
the circledpixel.
TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS 315

Figure 8. Imagesof upslopearea for a portion of the Walnut Gulch ExperimentalWatershedDEM. A


logarithmicscaleof gray shadesis usedwith lighter shadescorrespondingto higher values.This is a U.S.
GeologicalSurveyDEM with 30-m resolution.(a) Single-direction, D8, procedure;(b) Quinn et al. [1991]
procedure,MS; and (c) new procedure,D.
316 TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS

(c)

Figure 8. (continued)

alignedwith the grid axes,cardinal or diagonal,the Doopro- contours,as expected.DEMON is similarwith somespillover
cedure gives the same results as D8, and both are correct. into adjacentpixelsdue to the two-dimensionalflow represen-
However,when the topographicslopeis not alignedwith one tation. Doohas dependencefrom a narrower band 45 wide
of the grid directions,the proceduresdiffer. D8 introducesno upslopeof the point underconsideration. The axisof thisband
dispersion,but at the expenseof grid bias. Doofollows the is perpendicularto the contours.
topographicslopeat the costof introducingsomedispersion. For eachof theseexamplesurfaces,cone,inwardcone,and
Figure 6 showsupslopearea and the influencefunctionsfor plane, the true upslopearea wascomputedat eachgrid point
a portionof an inwardflowingconesurface.Figure7 showsthe and comparedto resultsfrom each of the DEM procedures.
dependencemapsfor a planesurfacenot alignedwith the grid. Table 2 presentsthe differencesbetweenthe theoreticalresult
The dependencemap reflectsthe fractionof a pixel'sarea that and each DEM procedure.
drainsto the designatedpixel. It servesto demarcatea zone of In evaluatingthe error statisticsthe grid directionbias in-
contribution,with shadingto denote the degree of contribu- troducedby D8, clearlyevidentin the figures,is responsiblefor
tion. On a planar surface the dependencemaps should be the large mean squareerror (MSE) on the cone surfaces.
straightlines perpendicularto the gradient.The D8 method Curiously,D8 doeswell for the plane becausethe area is the
gives straightlines following grid lines. MS has dependence samewhether one countsalong the grid or perpendicularto
from a broadarea,illustratingstrikinglythe problemswith MS, contours.This would not havebeen the casehad the ridgenot
even for a simple surface.The dependencemap for Lea's been alignedwith the grid. Quinn et al.'s [1991] MS method
[1992]methodis a stairsteppath roughlyperpendicularto the does best for the inward cone where the concave surface limits

Table 2. DifferencesBetweenTheoreticaland DEM-Computed UpslopeArea for Test ExamplesExpressedin Terms of


the Mean Error and Mean SquareError
Outward Cone Inward Cone Plane

Bias MSE Bias MSE Bias MSE


Mean(.4 -z) Mean((.4 -z) 2) Mean(,4 -z) Mean((.4 -z) 2) Mean(.4 -,) Mean((A -z) 2)
D8 -0.13 2.13 1.76 118.88 -0.17 0.065
MS -0.81 0.69 - 1.07 5.70 - 1.37 2.065
Lea's [1992] method - 1.29 2.41 -4.05 44.00 -2.57 7.912
DEMON -0.37 0.17 -0.37 19.23 -0.40 0.161
Doo -0.13 0.20 1.87 30.58 -0.17 0.065

Bias,meanerror;MSE,meansquare
error;
A, trueupslope
area;
,, computed
upslope
area.
TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS 317

Figure 9. Imagesof upslopearea for a portion of the TennesseeValley studyarea DEM, Marin County,
California.A logarithmicscaleof grayshadesis usedwith lighter shadescorrespondingto highervalues.This
is a 2-m resolutionDEM generatedfrom low-altitudestereoaerial photographs[Dietrichet al., 1992, 1993;
Montgomeryand Foufoula-Georgiou, 1993]. (a) Single-direction,D8, procedure;(b) Quinn et al. [1991]
procedure,MS; and (c) new procedure,D.

dispersion.However, it doespoorlyfor the plane and outward a bias toward overestimatingupslopearea by countinga con-
cone owingto its substantiallateral dispersion.For example, tributionof 1 evenif a flow path from a pixel onlyjust crosses
with the outwardcone,MS pixelsalongthe outwardedgehave the corner of a pixel. In terms of statisticsDEMON and
a componentof slopeparallel to the edgetowardsthe corners. are quite competitive.Both have small MSE and bias for the
Therefore some fraction of flow that should exit the domain at outward cone and plane and small MSE relative to Lea's and
the center is directed towardsthe corners,increasingthe sum D8 for the inward cone. DEMON is better on the inward cone.
of upslopearea aroundthe edgeabovethe area of the domain. This is due to the continuityacrosspixelsmaintainedby DE-
Lea's [1992]methodin all caseshaslargeMSE and biasdue to MON with its flow tube procedure.Toward the middle of the
318 TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS

Figure 9. (continued)

coneDEMON hasmanyflow tubesat spacingmuchfiner than catchmentarea by the methodsD8, MS, and Dc are shown.
the cell size.Dc mixesthe flow in each cell resultingin some Theseindicatesignificantdifferencesin the numberof pixels
blockinessevident in Figure 6e. Though informative, these computedto have a certain specificcatchmentarea at the low
statisticsdo not provide the completepicture. The good per- end of the specificcatchmentarea scale.At the high end the
formance of MS on the inward cone is due to its dispersion methods all give similar results. Specificcatchmentarea is
being contained and considerable concentric averaging
smoothinggrid effectsthat are not so smoothedby the other
methods.However, this averagingwhich contributesto small
error statistics
for a symmetrictestcaseis in generalundesir-
able as it results in loss of resolution.
Figure 8 showsthe upslopearea for portion of the Walnut
A.
Walnut
Gulch D8
MS

Gulch ExperimentalWatershedcalculatedusingD8, MS, and


Doo.Thisis a 30-mresolutionDEM from the U.S. Geological
Survey
HayMountain andTombstone7.5-min
quadrangles.
Thedifferences
atthisscale
arequite
subtle,
butit ispossible
to see more streakinessassociatedwith the grid from the D8
procedure(Figure8a) thanwith that from MS (Figure8b) and
Dc (Figure 8c). 1.5 2.0 2.5 3.0 3.5 4.0
Figure 9 showsthe upslopearea for a portion of the Ten- log(Specific
catchmentaream)
nesseeValley area near San Francisco,California. This is a
high-resolution(2-m grid) DEM generatedfrom low-altitude
stereoaerialphotographs[Dietrichet al., 1992,1993;Montgom-
.
.

..
B. TennesseeValley
z\
eryand Foufoula-Georgiou, 1993].At this scalethe differences : \ ' 138
aremuchmorenoticeable
thanwasthecaseforFigure8.The /'.:",,,,,
\,.. '.........MS
D8 procedureresultsin streaksalignedwith the grid indicating
that flow down hillsidesis biasedby the grid alignment.The
Doo
and
MSprocedures
donot
suffer
fromthese
problems.
The / .'
'".k,.
\
".
',.,
D resultsseemtohavemoresharpness
thanMS.Thisisdue / ,,. \

tolessgriddispersion.
Upslopeareas
werenotcomputedwith
DEMON or Lea's [1992]methodfor the real DEMs, because .........

in the case of Lea's method it performed poorly in the test


1 2 3 4
exampleevaluation(Table 2), and the DEMON codedoesnot
log(Specific
catchmentaream)
work automaticallygiventhe flat areasand situationsrequiring
exceptionspresentin this data. The DEMON downslopecode Figure 10. Histogramsindicatingthe distributionof specific
identified 6607 sinksin the Walnut Gulch DEM, even after the
catchmentarea by eachmethod.(a) Portionof Walnut Gulch
pits had been filled. Watershedshownin Figure8. (b) Portionof TennesseeValley
In Figure10,histograms
showing
thedistributioh
of specific area shownin Figure 9.
TARBOTON: DETERMINING FLOW DIRECTIONS AND UPSLOPE AREAS 319

small on hillslopesand large in valleys and on the channel Fairfield, J., and P. Leymarie, Drainage networksfrom grid digital
networks.Therefore, if the intent is demarcationof valleysor elevationmodels,WaterResour.Res.,27(5), 709-717, 1991.
Freeman,T. G., Calculatingcatchmentarea with divergentflow based
channels,for example,usingthe notion of a critical support on a regulargrid, Comput.Geosci.,17(3), 413-422, 1991.
area, the different methodswill give similar results.However, Jenson,S. K., Applicationsof hydrologicinformation automatically
if the intent is to use the specificcatchmentareasfor distrib- extractedfrom digital elevationmodels,Hydrol.Process.,5(1), 31-
uted modelingof hydrologicprocesses, suchasrunoffgeneration 44, 1991.
Jenson,S. K., and J. O. Domingue, Extractingtopographicstructure
or erosion,then the differentmethodswill givedifferingresults.
from digital elevationdata for geographicinformationsystemanal-
ysis,Photogramm. Eng.RemoteSens.,54(11), 1593-1600,1988.
Conclusions Lammers,R. B., and L. E. Band, Automatingobjectrepresentationof
drainagebasins,Comput.Geosci.,16(6), 787-810, 1990.
A new procedurefor the representationof flow directions Lea, N. L., An aspectdrivenkinematicrouting algorithm,in Overland
and the calculationof upslope area using grid-baseddigital Flow:HydraulicsandErosionMechanics,editedby A. J. Parsonsand
A.D. Abrahams,Chapman& Hall, New York, 1992.
elevationmodelswas presented.The procedureis basedon Mark, D. M., Network models in geomorphology,in Modelling in
representingflowdirectionasa singleangletakenasthe steep- Geomorphological Systems, chap.4, editedby M. G. Anderson,pp.
est downwardslopeon the eight triangularfacetscenteredat 73-97, John Wiley, New York, 1988.
each pixel. Results from the procedure were compared to Marks, D., J. Dozier, and J. Frew, Automated basin delineation from
digital elevationdata, Geo. Processing, 2, 299-311, 1984.
other grid-basedproceduresfor calculationof upslopearea
Martz, L. W., and J. Garbrecht, Numerical definition of drainage
from grid DEMs, using test examplesand low- and high- network and subcatchmentareas from digital elevation models,
resolution DEM data. On the basis of the evaluation of test Comput.Geosci.,18(6), 747-761, 1992.
statisticsand examinationof influenceand dependencemaps, Montgomery,D. R., and W. E. Dietrich, A physicallybasedmodel for
the new procedure performs better than D8, Lea's [1992] the topographiccontrol on shallowlandsliding,WaterResour.Res.,
30(4), 1153-1171,1994.
method,and MS and is comparableto DEMON [Costa-Cabral Montgomery, D. R., and E. Foufoula-Georgiou,Channel network
andturges, 1994]. Thenewprocedure overcomes theproblems sourcerepresentationusingdigital elevationmodels,WaterResour.
of loopsand inconsistencies that plagueplane-fittingmethods Res.,29(12), 3925-3934,1993.
such as DEMON. In real DEMs, significantdifferencesbe- Moore, I.D., R. B. Grayson,and A. R. Ladson,Digital terrain mod-
elling:A review of hydrological,geomorphological, and biological
tween the distributionof specificcatchmentarea were ob-
applications, Hydrol.Process., 5(1), 3-30, 1991.
tained dependingon the method used.Differencesare more Morris, D. G., and R. G. Heerdegen,Automaticallydrainedcatchment
markedasdigitalelevationdataresolutionincreases, especially boundariesand channelnetworksand their hydrologicalapplica-
at hillslopescales.Where resultsfrom the different methods tions, Geomophology, 1, 131-141, 1988.
differ, the choice of methodsbecomesimportant, and this O'Callaghan, J. F., and D. M. Mark, The extractionof drainagenet-
works from digital elevation data, Comput. Vision GraphicsImage
paper has arguedthat the new method presentedprovidesa Process.,28, 328-344, 1984.
simpleeffectiveapproachthat shouldwarrant consideration. Quinn, P., K. Beven,P. Chevallier,and O. Planchon,The predictionof
hillslopeflow pathsfor distributedhydrologicalmodelingusingdig-
ital terrain models,Hydrol. Proc., 5, 59-80, 1991.
Availability Quinn, P. F., K. J. Beven,and R. Lamb,The In (a/tan/3) index:How
to calculateit and how to use it within the Topmodel framework,
Softwarethat implementstheseproceduresis availableelec- Hydrol. Process.,9, 161-182, 1995.
tronically on the INTERNET from the author (dtarb@ Tarboton, D. G., The analysisof river basinsand channel networks
cc.usu.edu, http://www.engineering.usu.edu/dtarb/) and anony- usingdigitalterrain data, Sc.D. thesis,Dep. of Civ. Eng., Mass.Inst.
mousftp from fox.cee.usu.edu. of Technol., Cambridge,1989. (Also availableas Tech.Rep. 326,
Ralph M. ParsonsLab. for Water Resour.and Hydrodyn.,Dep. of
Civ. Eng., Mass.Inst. of Technol.,Cambridge,1989).
Acknowledgments.This work was supportedby National Science Tarboton, D. G., R. L. Bras, and I. Rodriguez-Iturbe, The fractal
Foundationgrant EAR-9318977. Thank you, Bill Dietrich and David natureof rivernetworks,WaterResour.Res.,24(8), 1317-1322,1988.
Montgomery,for use of the TennesseeValley DEM dataset.The Tarboton, D. G., R. L. Bras, and I. Rodriguez-Iturbe,On the extrac-
DEMON codewas suppliedby Mariza Costa-Cabral.I would like to tion of channel networksfrom digital elevation data, Hydrol. Pro-
thank Mariza Costa-Cabral,David Montgomery, and anonymousre- cess.,5(1), 81-100, 1991.
viewersfor their helpful reviews. Tarboton, D. G., R. L. Bras,and I. Rodriguez-Iturbe,A physicalbasis
for drainagedensity,Geomorphology, 5(1/2), 59-76, 1992.
Wolock,D. M., and G. J. McCabe, Comparisonof singleand multiple
References flow direction algorithmsfor computingtopographicparameters,
WaterResour.Res.,31(5), 1315-1324,1995.
Band,L. E., Topographicpartitionof watersheds with digitalelevation Wood, E., F. M. Sivapalan,and K. Beven, Similarity and scale in
models,WaterResour.Res.,22(1), 15-24, 1986. catchmentstormresponse,Rev. Geophys., 28(1), 1-18, 1990.
Beven,K. J., and M. J. Kirkby,A physicallybasedvariablecontributing Wu, W., Distributedslopestabilityanalysisin steep,forestedbasins,
areamodelof basinhydrology, Hydrol.Sci.Bull.,24(1), 43-69, 1979. Ph.D. thesis,WatershedSci., Utah State Univ., Logan, 1993.
Beven, K., M. J. Kirkby, N. Schofield,and A. F. Tagg, Testing a Zhang, W., and D. R. Montgomery,Digital elevationmodel grid size,
physicallybasedflood forcastingmodel (TOPMODEL) for three landscaperepresentation,and hydrologicsimulations,WaterResour.
UK catchments,J. Hydrol., 69, 119-143, 1984. Res.,30(4), 1019-1028, 1994.
Costa-Cabral,M., and S. J. Burges,Digital elevationmodel networks
(DEMON): A model of flow over hillslopesfor computationof D. G. Tarboton, Department of Civil and EnvironmentalEngineer-
contributingand dispersalareas,WaterResour.Res.,30(6), 1681- ing, Utah State University, Logan, UT 84322-4110. (e-mail:
1692, 1994.
dtarb@cc.usu.edu)
Dietrich, W. E., C. J. Wilson, D. R. Montgomery,J. McKean, and
R. Bauer,Erosionthresholdsandland surfacemorphology,Geology,
20, 675-679, 1992.
Dietrich, W. E., C. J. Wilson, D. R. Montgomery, and J. McKean,
Analysisof erosion thresholds,channel networks,and landscape (ReceivedJanuary16, 1996;revisedSeptember30, 1996;
morphology usinga digitalterrainmodel,J. Geol.,101,259-278, 1993. acceptedOctober9, 1996.)

You might also like