JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 99, NO.

Bll, PAGES 21,803-21,824, NOVEMBER 10, 1994

Bathymetric prediction from dense satellite altimetry

and sparse shipboard bathymetry
Walter H. F. Smith
NOAA Geosciences
Lab, NationalOceanService,Silver Spring,Maryland

David T. Sandwell
ScrippsInstitutionof Oceanography,
La Jolla,California

Abstract.Thesouthern
oceans
(southof 30S)aredensel)7
covered
withsatellite-derived
grav-
ity data(trackspacing2-4 km) andsparsely coveredwith shipboard depthsoundings (hundreds of
kilometersbetweentracksin someareas).Flexuralisostaticcompensation theorysuggests that
bathymetryanddownwardcontinuedgravitydatamay showlinearcorrelationin a bandof wave-
lengths15-160 km, if sedimentcoveris thin and seafloorrelief is moderate.At shorterwave-
lengths,the gravityfield is insensitiveto seafloortopographybecauseof upwardcontinuation
from the seafloorto the seasurface;at longerwavelengths, isostaticcompensation cancelsout
mostof the gravityfield due to the seafloortopography.We combinethis theorywith Wiener
optimizationtheoryandempiricalevidencefor gravitynoise-m-signal ratiosto designlow-pass
andband-passfiltersto usein predictingbathymetryfrom gravity. The predictioncombines
longwavelengths(>160 km) from low-pass-filtered soundingswith an intermediate-wavelength
solutionobtainedfrom multiplyingdownwardcontinued,band-pass-filtered (15-160 km) gravity
databy a scalingfactorS. S is empiricallydeterminedfrom thecorrelationbetweengravitydata
andexistingsoundings in the 15-160 km bandby robustregression andvariesat long wave-
lengths. We find that areaswith lessthan 200 rn of sedimentcovershowcorrelationbetween
gravityandbathymetrysignificantat the 99% level, and$ may be relatedto the densityof
seafloormaterialsin theseareas. The predictionhasa horizontalresolutionlimit of 5-10 km in
positionand is within 100 m of actualsoundingsat 50% of grid pointsand within 240 rn at
80% of these. In areasof very ruggedtopographythepredictionunderestimates the peak ampli-
tudesof seafloorfeatures.Imagesof the predictionrevealmanytectonicfeaturesnot seenon any
existingbathymetriccharts.Becausethe predictionrelieson the gravityfield at wavelengths
<160 krn, it is insensitiveto errorsin the navigationof soundinglines but also cannot
completelyreproducethem. Thereforeit may be usedto locatetectonicfeaturesbut shouldnot
be usedto assesshazardsto navigation.The predictionis availablefrom the National
GeophysicalData Centerin bothdigital andprintedform.

Introduction a marine gravity field (top), and the tracks of available

shipboard depth soundings [Smith, 1993; National
Topographic elevations are a fundamental physical
Geophysical Data Center, 1993] (middle). Also shown
characteristic of a planet, and yet the topographiesof Mars
(bottom)are local plate boundariesand a port (EasterIsland);
[Carr et al., 1977] and Venus [Ford and Pettengill, 1992] are
the ship coverageis biasedtoward portsand tectonicfeatures,
better mappedthan that of most oceanicareasof Earth [Smith,
while the satellite coverageis uniform.
1993]. In the SouthernOcean there are many gapsof hundreds
Marine gravity anomaliesand seafloor topographycan be
of kilometers between available bathymetric survey tracks
highly correlatedover a band of wavelengths[McNutt, 1979;
(Figure 1), and many of the data are of poor quality. In some
Smith and Sandwell, 1992; Neumann et al., 1993] so that
areas, the majority of available data were collected with only
ocean depths may be inferred from sea surfacegravimetry.
celestial navigation and discrepanciesin depths reported at
This was first suggestedin the 19th century by Siemens
intersecting survey tracks exceed 100-250 m at half of the
[1876], but it was more than a centurylater when Dixon et al.
track crossings[Smith, 1993]. The area southof 30Shas the
[1983] demonstrated that the techniquewas feasible. Dixon et
least detailed bathymetric coverage but the most detailed
al. usedprofiles of geoidheight data obtainedalong tracksof
gravity coverage thanks to recent declassificationof Geosat
Geodetic Mission (GM) altimeter data [Marks et al., 1993].
the Seasatsatellitealtimeterin an area where the bathymetry
was already well known from conventional surveys. This
Figure 2 shows the ground tracks of the Seasat,Geosat, and
"ground truth" allowed Dixon et al. to demonstratethat the
ERS 1 satellitesusedby Sandwelland Smith [1992] to compile
predictions dependedstrongly(+1 km of topography)on the
assumedform of isostaticcompensation,and that topographic
Copyright
1994bytheAmerican
Geophysical
Union. featuresout of the profile plane also had a stronginfluenceon
Papernumber94JB00988. the solution. Subsequentstudies[White et al., 1983;Baudry et
0148-0227/94/94JB-00988505.00 al., 1987; Craig and Sandwell, 1988; Jung and Vogt, 1992;

21,803
21,804 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

Goodwillie and Watts, 1993] have also been confined to the bp(x) = d(x) + S(x)g(x), ()
vertical plane under altimeter profiles and have thus been
forcedto makesomeassumptions aboutsymmetrybecauseof wherethe passbandpredictionis the productof the band-pass-
their lack of gravity data in the across-trackdirection. Now filtered and downwardcontinuedgravity g(x) and the scaling
that completegravity coveragein two horizontaldimensions factor S(x). Notethatwiththismethod, bp(x)willnotfit the
is availablefrom the GeosatGM data,it is time to attempt original soundingsexactly. Soundingswhich are misplaced
bathymetricpredictionthroughoutthe southernoceans. by a few tens of kilometers will introducelittle error into d(x)
In this paper,we presenta new techniquewhichmakesfull but would have a severeeffect at shorterwavelengths,where
use of the GeosatGM data, avoidsmost assumptions about we chooseto rely on the gravity data to locate the predicted
isostaticcompensation, and allowsfor regionalvariationsin features.
the correlationbetweenbathymetryand gravity. We show An example is shownin Plate 1. The top image showsthe
that these regional changesare necessaryto accommodate sea level gravity field of Sandwell and Smith [ 1992], while the
changesin the thicknessof sedimentson the oceanfloor. The middle image showsthe predictedbathymetry. For compari-
images of the seafloor we obtain resolve tectonicdetails to 5- son, the bottom image shows the best available gridded
10 km in locationand within 100-250 m in depthin many bathymetrydata known as ETOPO-5 [National Geophysical
cases. However,the methodunderestimates peak amplitudes Data Center, 1988], which is based on the earlier DBDB-5, or
in ruggedareasand at very tall seamounts,so it is not suitable SYNBAPS data [Van Wyckhouse, 1973] (ETOPO-5, Earth
for charting hazards to navigation. In the next sectionwe TOPOgraphy-5 minute; DBDB-5, Digital BathymetricData
give a broad overview of the technique;later sectionswill Base-5 minute;SYNBAPS, SYNthetic BAthymetricProfiling
explaineach stepmore fully. System). The prediction shares some features with each of
these other images, and one could imagine combiningthe
General Description regional features of ETOPO-5 with the local features of the
gravity to produce the predicted bathymetry. If S(x) were
The relationshipbetween seafloor topographyand sea everywhere constant and ETOPO-5 matched d(x)exactly at
surfacegravity is convenientlydescribedby functionswhich long wavelengths, then the middle image (the prediction)
dependon the wavelengthof the topography[Dorman and would be a simple combinationof the top and bottom images.
Lewis, 1970; McKenzie and Bowin, 1976]. Theoretical and In fact, ETOPO-5 does not fit the ship data very well [Smith
empiricalstudies(discussed
below)suggestthat we canhope and Wessel, 1990; Smith, 1993], and variations in S(x) make
for correlationbetweengravity and bathymetryonly in a the predictionmore complicated,and more interesting. We
limited band of wavelengthsbetweenabout 15 and 160 km. find that S varies geographicallyin accordancewith the thick-
The shorter-wavelengthpart of the gravity field (<15 km) is nessof sedimentcover, and in someareas,gravity and topog-
attenuatedby upward continuationfrom the seafloor to the sea raphy are uncorrelatedand S = 0. Estimation of S at the areas
surfaceso the signal-to-noiseratio is small, and it is unsafeto labeled A, B, and C in Plate 1 will be discussed in the section
attemptprediction. Similarly, at longer wavelengths(>160 on the inverseNettleton procedure.
km), seafloortopography can be isostatically
compensated
so
that it produceslittle or no gravityanomaly.Our approachis
to use gravity to predict bathymetryin the 15-160 km band.
Data Preparation
Shorter-wavelength topography(<15 km) can only be recov- Digital shipboardsoundingssouthof 30S were derived
ered throughdetailed shipboardsurveys,while longer-wave- from two sources.The primarysourcewasassembledfrom the
length (> 160 km) topography may be constrainedby low- Lamont-Doherty
EarthObservatory
holdingsandput through
pass-filteringgrids of existing ship soundings. extensivequality controlprocedures[Smith, 1993]. This was
Within the prediction band, the ratio of topography to augmented by datacompiledby the National GeophysicalData
gravity, S(x), is determined empirically in areas where these Center [1993]. Most of these additional data were recent sur-
are well correlated. The topography-to-gravity ratio is veysalongthe Antarcticcoastby foreignAntarcticagencies,
allowed to vary at long wavelengths (> 160 km) to particularlythe British and Japanese;a few were newly
accommodateregional variations in sediment thicknessand contributeddatafrom domesticlaboratories.Notableamong
seafloor density. High topography-to-gravitycorrelationsare these were a R/V Moana Wave survey of the Australian-
generally found over unsedimentedtopographybecauseof the AntarcticDiscordanceandR/V ThomasWashington surveys
of
large seawater-to-rockdensitycontrast. Low correlationsare the EastPacificRise andthe LouisvilleRidge. All new data
generally found in areas where the topographyis buried by from the CD-ROM were put throughthe qualityscreening
thick sediments;in these areas the buried structuresproduce recommended
by Smith [1993]. The Japanese
datawere found
gravity anomalies, while the seafloor relief is low. to havenavigationreportingerrorswhenlatitudeor longitude
Fortunately, these sedimented areas are nearly flat at short wasequalto an integernumberof degrees;theseerroneous
data
wavelengthsso that we may take S=0 here. In practice, we weredeleted.A totalof 581legswereusedin ourstudy(Figure
band-passfilter the gridded ship soundingsand gravity and 1).
downwardcontinuethe gravity to the regional seafloordepth. Thesedataweregriddedontoa meshhaving20 samples per
The theory predicts a linear relationship between these data,
degreelongitudeand40 samplesper degreelatitudeandspan-
and so we use a linear regressionmethod to estimate both ning an area between25.8Sand 73.8S. This procedure
topography-to-gravityratio and correlation. This method, allowedus to useresultsin the area30-70Swithoutedge
which we call the "inverse Nettleton procedure," uses only effectsfrom our filters and resultedin grid "cells"whichare
those points in the grid where actual soundingsare available.
squares approximately 2.75 km on a sideat 60S.The gridwas
The total predicted bathymetry bp(X)is thesumof the preparedby first assigningall soundingsto their nearestgrid
passbandprediction and long-wavelengthregional depth d(x): pointandthenat eachgridpointwith oneor moresoundings
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,805

120 130 140 150 160 170 180 190 200

i

! -50
I

-55! I
I
!
I

i-6o
I

.65
.i..
! _65

-50 I '50'
!
!
-- I
!
I
I

i-6o
I

x,
'"I -65:
.50 - .50 -

-55:

.60

I
.65

120 ' 130 140 150 160 170 180 190 200 *

Plate 1 (top) Free-air gravity field from Sandwell and Smith [1992]. Areas labeledA, B, and C are discussed
in connectionwith Figure 6, and the dashedline is the locationof the profile shownin Figure8. (middle)
Bathymetric predictionmade from gravity by the methodof this paper. (bottom) ETOPO-5 bathymetry for
comparison.
21,806 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

in the bathymetricpredictionprocedureoutlined above. Here

0 o we give a detailed descriptionof the filter design and imple-
mentation. The theory is developed in terms of the Fourier
transforms of gravity and topography data. A flat Earth
approximation is used throughout the analysis because the
o longestcutoff wavelengthin any filter is only 160 km, which
is much less than the radius of Earth. Our grids are sampled
with equal spacingin latitude and longitude, so the distance
betweenadjacenteast-westgrid pointsvarieswith latitude. To
accommodate this latitude-dependentscale, we divided the
30-70 latitude range into smaller overlapping strips, cosine
windowed each strip, performed the filtering by fast Fourier
transform with the correct scale in each strip, and then
combinedthe strips. All filtering and downwardcontinuation
describedhere was performedin this fashion.
o

The designof the bathymetricpredictionfilter is motivated

by theoretical and empirical studies of Z(k), known as the

225 230 235 240 245 250 255

_25 _25

180

_30 _30
Figure 1. Ship tracks of digitally available bathymetric
soundings.One-quarterof Earth'ssurfaceis shown(southpole
to 30S)in an equal-areaprojection.
_35 _35

assigned,retaining the median of the assignedvalues. Grid

points so constrainedwere marked for use later in the "inverse
Nettleton" determination of the topography-to-gravityratio. _40 _40
Grid points without control data were interpolated using a -25 .............. -25
continuous curvature spline in tension [Smith and Wessel,
1990]with a tensionv/[lae'of 1, corresponding
to a harmonic
spline. With this approach, the grid points constrainedby
_30 _30
ship soundings retain their initial values, and the harmonic
spline assuresthat the solution has local maxima and minima
only at these points [Smith and Wessel, 1990]. The resulting
grid has hills and valleys only where the ship data require _35 _35
them. During this griddingprocess,areason land were fixed to
valuesinterpolatedfrom ETOPO-5.
The gravity data set [Sandwelland Smith, 1992] was derived
from radar altimeter measurementsof sea surfaceheightsmade _40o
Ix ]X, ,.__
by the Seasat, Geosat, and ERS 1 satellites [Sandwell, 1984, _25 _25
1992]. The precision of the Seasat, Geosat, and ERS 1
altimeter data has been studied by Marks and Sailor [1986],
Sandwell and McAdoo [1990], and Sailor and Driscoll [1993], _30 _30
respectively. The resolutionof the derived gravity dependson
the spacingof the satellite tracks (2-4 km, Figure 2, top) and
also to a limited extent on latitude; the estimation method used
by Sandwell and Smith [1992] takes account of this depen- _35 _35

dence and filters the data appropriately. Therefore in the

appendix we assume that the noise spectrum of the gravity
map is isotropic, i.e., independent of azimuth. The gravity
_40 _40
grid has a root-mean-squareerror of 3-5 mGal, high correla-
225 230 235 240 245 250 255
tion with ship gravity measurementsat wavelengthslonger
than 20 km [Smith et al., 1993], and high correlation with Figure 2. Mercatortrack mapsfor an areaof the SouthPacific
ship bathymetry for wavelengths greater than 26 km 2790 km E-W by 1670 km N-S. (top) Seasat,ERS 1, and
[Neumann et al., 1993]. The gravity grid was resampledonto Geosat satellite tracks. Geosat GM data available south of
the samegrid as the initial bathymetrygrid. 300Sare spaced2-4 km apart;north of 30Sthere are 40 by
100 km gaps in coverage. (middle) Ship soundinglines in
same area. There are 500 by 750 km gaps, and coverageis
Admittance Theory
biasedtowardports(EasterIsland) and particularfeatures(East
Flexural isostaticcompensationtheory is usedas a guide to Pacific Rise). (bottom) Geographic references. El, Easter
design various low-pass,high-pass,and band-passfilters used Island; EPR, East Pacific Rise; JF, Juan Fernandezmicroplate.
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,807

gravitational"admittance"
in the literature[Lewisand Dorman wavelength(km)
1970; Dorman and Lewis, 1970; McKenzie and Bowin, 1976; oo 100 50 33.3 25 20
Watts, 1978, 1979; McNutt, 1979]. Z(k) is the transfer 100
function of a linear, isotropic, and spatially invariant filter
which takes a topographyfield h(x) as input and producesa
gravityfield g(x) as output. In the Fouriertransformdomain
this is written 2rcFp

= (2)

The definitions of Bracewell [1978] are used here for two-

dimensional Fourier and Hankel transforms: H(k) is the
Fouriertransform
of h(x), wherex is [x,y]andk is [u=l/)x,
v=l/)y]. A one-dimensional
Hankeltransform
is usedfor
isotropicfunctionsdependentonly on the scalardistancer =
Ixl, and scalar wavenumberk = Ikl; Z(k) is the Hankel trans-
form of z(r). In formulating (2), the topographyh(x) is
defined as the local relief of the seafloor measured from a
25-
I =135km
'
constant
regionaldepthd. Parker [1973]hasshownthat(2) is
only approximately
true, as the correctrelationship
is nonlin-
0
ear. The approximationis a good one when h is small
comparedwith d. We retain this linear approximation
throughoutthe paperanddiscussits limitationsin evaluating
the bathymetric prediction.
In flexural isostatic compensationtheory, the seafloor
topographyh actsas a loadon a thin elasticlithosphere, and ! Ii X=135km /
the lithosphereflexes under the load in a linear, spatially
invariant,and isotropicmanner. The gravityfield g(x) is the
40- I '=800km
sum of the direct topographic effect and its isostatic
compensation. Whenthe wavelength of the topographic load
is much less than the flexural wavelengthof the lithosphere
(defined below), the admittancefunction is given by the
uncompensated
model

z (k): 2rcFp
exp[-2rckd], (3)
20
where d is the regional depth, p is the densityof the seafloor
relative to seawater, and F is the Newtonian gravitational
constant (2rcFp is known as the Bouguer constant). The
exponentialdecay with increasingwavenumberis known as
"upwardcontinuation"and is shownin Figure 3 (solid curve). 10
0.00
....................
0.01
.... .....
...........
0.02 0.03 0.04 0.05
The gravity anomaly producedby an uncompensated conical
seamount(2 km high and 60 km in diameter) is shown in k (1/km)
Figure 4 (long dashedcurve). The exponentialdecayin Z acts
Figure 3. Transfer functionsversuswavenumber. Scale at
as a smoothingfilter, so that the sharpedgesin the topogra-
bottom in wave,umber; scale at top in equivalentwavelength.
phy are not present in the gravity profile. Also, the peak
(top) Z, the transferfunctionfor topographyinput and gravity
amplitudeof the gravity anomalyis lessthan 2rcFptimesthe
output. A constantvalue (dotted line) holds for uncompen-
peak amplitudeof the topography. Implicit in the bathymet-
satedtopographyon the gravity observationlevel. Upward
ric prediction problem is the need to undo this smoothing,
which is unstable.
continuationto a level d producesexponential decay (solid
curve). Isostatic compensationcausesZ to approachzero at
When the wavelengthof the topographicload matchesor
wavelengths longerthana parameter) (dashedanddot-dashed
exceedsthe flexural wavelength,the plate deflectsunder the
curves). (bottom)Q = Z-1, the transferfunctionfor gravity
load causinga downwarpingof the Mohorovicicdiscontinuity
input and topographyoutput. Each curveis the reciprocalof
(Moho), m(x). Assumingthat the deflection of the plate is
the correspondingone in the top panel.
much less than the flexural wavelength and the plate has no
inelasticresponse,the Moho deflectionis relatedto the topo-
graphicload by anotherlinear isotropicfilter [Bankset al.,
1977]: q)(k)= [ 1+(,k)4]
-1. (5)

Here(Pc- Pw)is thedensity contrast

of theseafloor relativeto
(4) seawaterand (Prn- Pc) is the densitycontrastof the mantle
(pm-pc) relative to the crust. The flexural wavelength) is the wave-
lengthat which = 0.5; at wavelengths
muchlongerthan)
with the topographyis Airy compensated,while at wavelengths
21,808 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

I I I
lOO by the dotted curve in the top panel of Figure 4, and the
combined gravity is shown as the solid curve. The filter (6)
reducesthe peak gravity amplituderelative to the uncompen-
75- satedmodel and introducesnegativeside lobesso the resulting
gravity does not look exactly like the topography. Attempts
to recreate the missing signal with an inverse filter are
unstable.
50-

Window Carpentry and Filtering

25 By analogy with(2), we maydefinea linear,isotropic,
spa-
tially invariantprocesswhichtakesgravityas input and
producesseafloortopography
as output;it will havea transfer
function Q(k) suchthat
0

H(k) = G(k)Q(k}. (7)

-25 Inspection
of equations
(2) and(7) suggests
that Q(k)= Z-l(k)
[Dixon et al., 1983]. However, this definition results in a
functionwhich approaches
infinity at both the low and high
I I I wavenumber ends of the spectrum (Figure 3, bottom) and
whose Hankel transform does not exist. Thus topography
prediction using (7) is unstable. At short and long wave-
-2
Pw lengthsany noise in the gravity field, or slight deviationsin
density from the assumeddensity, will be greatly amplified
-4 resultingin a poor prediction. Moreoverat long wavelengths,
uncertaintyin the flexural wavelengthwill result in a poor
predictionbecauseQ is very sensitiveto ). Thus Q mustbe
-6
takenas a band-limited
approximation
of Z- [Dixonet al.,
19831
Pc
=z (a)
-10 tn
where W(k) is any band-passfilter (also called a spectral
window function) which stabilizesthe predictionproblemand
forces Q(k) to have a Hankel transform which exists. This
means that predictionsformed using equation (7) can only
resolvea limited range of wavelengthsin h.
-14 , ,
-200 -100 0 100 200
Dixon et al. [1983] used the above isostaticflexure theory
for Q to attempt a prediction. Their study covereda limited
km area so they were able to assumethat d and p were constants,as
the above theory requires. They chosea rectangularwindow
Figure 4. (top) Gravity anomaliesdue to a seamount.Dashed W(k), which producedspuriousoscillationsin their predicted
curve, effect of topographyonly; dotted curve, effect of com- bathymetry due to Gibbs' phenomenon [e.g., Bracewell,
pensationonly; solid curve, combinedeffect. (bottom) Cross 1978]. Their window includedsomeisostaticallycompensated
sectionthrough the ocean crust showinga conical seamount wavelengths in the passband,and so they found that their
(h, dashedcurve) and the moho deflection (rn, dotted curve) resultsdependedstronglyon the value of ) they assumed:
which compensates it. Also shownare water depthd, crustal predictedheights of seamountsvaried by 1 km accordingto
thickness
c, anddensities
Pw,Pc,Pmappearing
in thetext. the choiceof ). Finally,andperhapsmostimportantly,they
used only one-dimensional profiles of gravity data and
assumedthat both the gravity and topography fields were
lineated perpendicularto the track, when in fact their study
much shorterthan 3 the topographyis uncompensated.
The
admittance function for this model is
area (the Musician Seamounts) contains many pointlike
features.
The flexural wavelengthof the lithosphere,), is relatedto
z() = 2r0exp-:a]{ 1-exp-2kc]q(k}}
(6) the thicknessof the elastic part of lithospherewhich in turn
increases with the age of the lithosphere [Watts, 1978;
wherec is thenormal
crustal
thickness
(7 km)and p = Pc- Pw' Caldwell and Turcotte, 1979; Sandwell and Schubert, 1982]
Admittance functions for 3 of 800 km and 135 km are shown becausethe base of the elastic layer follows an isotherm. The
in Figure 3 (dashed/dotted
curves);theseare typicalminimum thinnestlithosphereis found at the seafloor spreadingridges.
and maximumvaluesfor ). The isostaticcompensation of the Over the fast spreadingEast Pacific Rise, elastic thicknesses
conical seamountis diagrammedin Figure 4 (bottom)for ) = range between 2 and 6 km, while over the slower spreading
135 km. The gravity effect of the Moho deflectionis shown Mid-Atlantic Ridge, the elastic plate is a somewhatthicker 7-
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,809

13 km [Cochran, 1979]. All of the ridges southof 30Sare W2 = 0.5whenk- = 15,20,and25km,forwater

depths
of2,
spreading
a ratesof 98 mm/yror less.We followWattset aL 4, and 6 km, respectively.
[1980] and assumean elasticthicknessof 5 km () - 135 km) The combined filter W(k)= W(k)W2(k) is shownin Figure
as a minimum value. Topographicfeatures on the seafloor 5 (top), and its Hankel transformw(r) in Figure 5 (bottom).
which form away from the ridge axis may exhibit admittance The dotted, solid, and dashedcurvescorrespondto water depths
functions indicating a range of elastic thicknesses[Watts, of 2, 4, and 6 km, respectively. The impulse responsew(r)
1978], and in some cases,quite different values may be found
within 150 km of one another [Smith et al., 1989]. If the
coolingof the lithospherefollows the plate model of Parsons wavelength(km)
and $clater [ 1977], then the maximumvalue of ) we would o 50 25 16.7 12.5 10

expect to find is about 800 km; this range of wavelengths

betweenapproximately135 and 800 km has been called the 1.0

"diagnosticwavebandof flexuralresponse"[Watts, 1983].

We designedour filter W(k) as the productof two filters
Wl(k) andW2(k)to suppress the long-andshort-wavelength
singularities,
respectively.
To designW1,we assumed that )
>= 135 km in the southern oceans, and we formed the function 0.5
Qoo/Q135,
wherethesubscript
indicates
the) valueand)= oo
corresponds
to the uncompensated
case(solid curvein Figure
3, bottom); this ratio removes the effects of d and p and
isolatesthe effect of ). From equation(6) it follows that

0.0
Qoo
= 1-exp [-21[kc]135(k), (9)
Q135 I I I I
\

wherec is thethickness
of thecrust.TheratioQoo/Q135
is well 3 ! \
approximatedby a simpler Gaussianhigh-passfilter with ! \
parameters = 30 km !
! \

2-
Wl(k)= 1-exp[-2(ks)2]. (lO)

W1= 0.5whenk-1= 160km. Thisfilterremoves

theflexural
wavebandand longer wavelengthsand is easier to computein
the form (10) than (9).
The low-passfilter W2 wasdesigned
to suppress
the high
wavenumber band where the exponential growth in Q due to
the downward continuationbecomeslarge. At wavelengths
i i i i

wherethe gravitydataare very accurate,W2 shouldbe near1 0.00 0.02 0.04 0.06 0.08 0.10
(the passband),while at wavelengths where the gravity data
areunreliable,W2 shouldbe near0 (thestopband).The choice k (1/km)
of transition wavelength is important; if it is too low, the
predicted topography will have low resolution, while if it is
too high, the predictionwill be overwhelmedby noise. If the 1.0

noise-to-signal ratio as a function of wavenumber R(k)is

known,then W2 canbe designed
so as to minimizethe mean
squareerror in the downwardcontinuedgravity field (and hence
the predicted topography). This method of optimizing filters
to minimize the variancein an estimatedsignal is from Wiener
0.5
[1949]. In the appendix we sketch the derivation of the
Wiener filter and show how published studies of the spectral
coherency in Geosat altimeter data [Sandwell and McAdoo,
1990] can be used to make an inspired guessat the functional
form of R(k). It turnsout that W2 dependson the regional
water depth d, becausethe gravity signal strengthdecaysat a o.o

rate dependingon d (Figure 3, top) while the noise processis

independent
of d. The formwe adopted
for W2 is
r (km)
W2(k)
= { 1+Ak4exp[4;kd]}
-1 (11)
Figure5. The filters (top) W(k) and(middle)W(k) exp[2nkd],
which can be found by substitutingequation (A13) into (A8). and (bottom) the impulse responsew(r). Dotted, solid, and
In (11), A is a constant chosen so that R(k) fits the observed dashedcurvescorrespondto water depthsof 2, 4, and 6 km,
spectral
coherency;
we usedA = 9500km4. Withthisvalue, respectively.
21,810 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

shows that the filter acts as a local averaging operation,which inverse of Nettleton's (hence "inverseNettleton procedure"),
limits the resolution of our prediction. One definition of but we use a robustregressiontechniqueconstrainedto pass
resolving power [Backus and Gilbert, 1968] concerns the through the origin.
averaging length of a linear system, which Constable and In order to obtain S(x) as a smoothlyvarying functionwith
Parker [1991] take as the width of the impulse responseat spectralcontentcomparableto d(x), we estimateS at points
half-maximum amplitude. For d = 2, 4, and 6 km this width is on a grid (the "Nettletongrid") sampled2.5 degreesin longi-
8.5, 11.4, and 13.8 km. Another measureof resolvingpower tude and 1.25 degreesin latitude. This yieldsa grid spacingof
is Rayleigh's criterion as used in optical imaging systems 135 km at 60S. At eachNettletongrid point we form the set
[Goodman, 1968]; by this criterion, two point featurescan be of all g,h pairs which are at pointsin the h grid originally
distinguishedif they are separatedby a distancegreater than constrainedby ship soundingsand which are within a 135 km
the distancefrom the center of w(r) to the first minimum. For d radius of the Nettleton grid point. Each pair is assigneda
= 2, 4, and 6 km this distance is 11.2, 14.6, and 19.2 km. weight basedon its distancefrom the estimationpoint usinga
Becausew(r) dependson the water depth,our methodwill have cosine window. If there are n data in the set then the sum of
higher resolution in shallow water. the weights will be about 0.3 n when the data are distributed
The theory for Z(k) and Q(k) is derived assuminga constant uniformly through the region r < 135 km. We found that
water depthd. We allow a slowly varyingwater depthd(x) by weight sumsof 10 or more gave good estimatesof S.
using the low-passfiltered versionof the observedbathymetry In Plate 1, three areas are labeled A, B, and C, and the data
grid bo(x)' from these areas used in the inverse Nettleton procedureare
shown in Figures 6a, 6b, and 6c, respectively. Area A is
D(k) = Bo(k)[ 1- Wl(k). (12) located on young seafloor along the Pacific-AntarcticRidge,
area B is on older seafloorin the vicinity of a large fracture
The local seafloor topography is the band-pass-filtered zone system,and area C is on the South Indian AbyssalPlain.
version of the observedbathymetric grid: Sedimentcover is negligible in area A and increasesfrom A to
B to C, reducing the spread in the distributions of g and h.
/-/(k) = Bo(k)W(k), (13) Figures 6a, 6b, and 6c show histogramsof g in the top panel,
h in the secondpanel, and the scatterplot of h versusg in the
where d is set to 0 in (11) as used in (13). bottom panel. The scatterplots are shownat the samescalein
Finally, the downward continued and band-pass-filtered each figure to illustrate the reductionin spreadcausedby the
gravityg(x) is relatedto theobserved gravitygo(X)by sedimentcover. The histogramschange bin width and scale
from one figure to the next in order to show the data
G(k) = Go(k)W(k)exp[2rkd. (14) distributions. The effect of increasing sediment cover is to
increasethe skewnessof the distributions,particularlyin the
This filter is shown in the middle panel of Figure 5. Since the h data.
depth varies spatially and (14) dependsstrongly on depth, a In simple linear regressionwe would fit a line to g,h pairs
different filter was applied to each grid cell. In practice, an by ordinary least squaresassumingthat g was known perfectly
initial solution was calculated using (14) for constant values and h contained some random error. We could take S as the
of d, in multiples of 1 km of depth. Then the actual g(x) was slope of that line. If there are errorsin both g and h, simple
obtained at each grid cell x using d(x) and linearly interpolat- linear regressionleads to biasedestimatesfor the slope of the
ing among the solutions found using constant d. In this regression line [Brownlee, 1965]; in this case a better esti-
manner,g(x) is the band-passedgravity field "drapedover" the
mateisS= q-'_Oh/Og , whereOg,oharethestandard deviations of
regional depth d(x). We are now assumingthat the theory g and h [Bendat and Piersol, 1986]. In either case the line
derived using a constant d can be applied locally (at wave- passesthrough the point with coordinates mean g, mean h,
lengths shorter than 160 km) since d(x) varies only region- and the portion of the variancein h which can be explainedby
ally (at wavelengthslonger than 160 km). its relationship with g may be characterized by the linear
correlationcoefficient. The parameterestimatesare maximum
Inverse Nettleton Procedure likelihood only if the data are drawn from a normal, or bivari-
ate normal, distribution [Brownlee, 1965].
After the above processing, we have all but one of the Because the data exhibit nonnormal distributions, we did
componentsneeded for the bathymetric prediction given by not use the conventional standard deviation in estimating o,
equation (1). Thetotalpredicted bathymetry bp(X)isthesum the spreadof the data; instead, we defined o by o = 1.4826 E,
of the passbandprediction S(x)g(x) and long-wavelength where E is the weighted median absolutevalue of g or h. The
regional depth d(x). Since the analysis is limited to the median absolute deviation is a well-known nonparametric
uncompensated band,the theoretical valueof S is [2xFp] -1, estimateof the spreadof a distribution,and the factor 1.4826
where p is the density of the seafloor material relative to makes this o equivalent to the o of a normal distribution(the
seawater. If g and h obey the theory, then h = Sg. Our task is expected value for E of a normal distribution is o/1.4826)
to determine whether g and h are in fact correlatedand the best [Rousseeuw and Leroy, 1987]. The weighting we applied in
value of S to predict h from g. Nettleton [1939] suggested the calculation of E makes our estimate of o depend more
that the appropriate density for Bouguer gravity reductions strongly on points near the center of the window. When the
could be found by trial and error, selectingthe density which mean value of a distribution is unknown, one computes the
yielded the Bouguer correctionlooking most like the free-air standard deviation with respect to the mean; similarly, E is
anomaly. In essence, Nettleton's method for density usuallycomputedas the medianabsolutedeviationof the data
estimationis a regressionprocedure: find p suchthat 2nFph from some location estimate. We computed E using the
best fits g by regressionof g onto 2rFh. Our problemis the median absolute value of the data, which is equivalent to
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,811

40 Like the ordinary linear correlation coefficient, the magnitude

of x is < 1, with the sign indicating the senseof the correla-
mGal bin
tion. The distribution of x is known in the case of the null
30 34.5 mGal hypothesis(Ho) thatg andh are uncorrelated,
and we can
calculate the confidence with which this hypothesis can be

I I I
30

1 regal bin
10

0.8 mGal
20
o
-lOO -50 0 50 100

mGal
10

40

30 rn bin

30 c = 341.3 m
-25 0 25

mGal
20
30

10 rn bin
10

J = 65.9m
20
0
-2000 - 1000 0 1000 2000

- I -
1TI
10

I I I
2OOO

ee

e_
-750 -500 -250 0 250 500 750
1000 -
9.9m/mGal %.....4[.. 1TI
0.60 ..... ,'.,
..,
.', 2OOO i I I

_
me

1000 S = 6.1 m/mGal

- 1ooo 0.36
e

-2OOO

- 100 -50 0 50 100

mGal
- 1ooo

Figure 6a. Histogramsand scatterplot of gravity and ship-

constrainedtopographyfrom area A in Plate 1, representing
young seafloor with little sediment.
-2000 i i I

assumingthe data are distributedaroundzero; this is so that if -100 -50 0 50 100

wetakeS= 't-IJh/IJg,
thenS willbetheslope
of a linethrough mGal
the origin.
We also used a nonparametricestimateof x, the correlation Figure 6b. Same as Figure 6a, exceptfor area B, representing
between g and h, called "Kendall's tau" [Press et al., 1986]. partial burial of basementby sediment.
21,812 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

30 [ [ I [ [ I [ dashedlinesareformedin the samemannerbut usingtheroot-

0.5 regal bin
weighted-mean-square to estimateo and the ordinarylinear
correlationcoefficientto determinethe sign of the slope;
5.3 mGal theserepresentthe weightedleast squaressolution. Areas A
20 and B illustratethat the Z methodyields a line which seemsto
fit the majority of the data more closely;the dashedline is
pulledtowardthe outlyingdata. In areaC, the signsof the

correlation coefficients are different, but both are so small in
10 magnitude
thattheconfidence
thatH0 canberejected
is also
small; thesedata show no significantcorrelation.
The confidence
with whichwe can rejectH0 andthe
estimatedvalue of x for all Nettletongrid pointsare shownin
the top two panelsof Figure 7. Someof thesecan be relatedto
-20 -15 -10 -5 0 5 10 15 20 the major geologicalfeaturesshownin the bottompanel of
Figure 7. Both of these are shown as zero where no data are
mGal
available. In most other areas, the confidence in the
correlation is greater than 99%. Correlation.coefficientsare
[ I i
30
particularly high over some continentalshelf breaks and some
1 robin active plate boundaries (Mid-Atlantic and Southwest Indian
Ridges;Chile, Hjort, and Tonga Trenches'ScotiaSea), as well
o=20.7m as the aseismicLouisvilleandBrokenRidges. All of theseare
20-
areasof largesignalsin g andh, similarto areaA in Figure6a.
Correlationsare low, and even slightly less than zero, and
confidencesalsolow, over the abyssalplains. Theseareasare
like areaC in Figure6c. Otherareasyield intermediatevalues.
10
At two Nettleton grid points southwestof the intersectionof
the Pacific-Antarctic ridge and the Eltanin Fracture Zone
system(220E, 55S) there is a stronglynegativecorrelation
with high confidence; this is, of course, contrary to
o- I theoretical expectations. We examined this area and found a
- 1oo -50 o 50 100
narrow ridge in both the gravity and depth data, but in slightly
1TI different locations. The depth soundingswere all collectedby
the R/V Verna using only celestial navigation, and we believe
2000 [ I I the soundingsare mislocatedhere, causingthe inverse correla-
tion.
These experiences with estimating correlations led us to
define S as follows: We only estimateS where the sum of the

1000 - -3.9 m/mGal weights

is at least10. We takeS = Oh/Ogwhenever
the
confidencethat H0 canbe rejectedis 95% or moreandx is
-0.04
positive. If x is negative or the rejection confidence is less
than95%, we proceed
basedon thevalueof oh. If Oh < 50 m,
we assume we have a situation like in area C, where the
seaflooris nearly flat and g is uncorrelatedwith h; we set S=0.
If Oh > 50 m, we assume
thattheseaflooris roughenoughthat
g and h would have a strong correlation if the ship data were
better navigated;we leave S unspecifiedat this point. We feed
-1000 - these values of S into the Smith and Wessel [1990] gridding
program used earlier for the ship data and use it to estimate
S(x) on the same grid mesh as the g and h grids. The result is
shownin Figure 7, third panel from top. If the seafloortopog-
-2000 I i i raphyhas a densityof 2600-2800kg/m3, appropriate
for
-11 0 -50 0 basalt, then we expect S in the range 13-16 m/mGal. Slopes
50 100

mGal in this range occur over someparts of the active ridge system.
Slopesare generally less than 6 m/mGal in the abyssalplains.
Figure 6c. Same as Figure 6a, except for area C, representing Intermediate values may indicate areas where large features
an abyssal plain. formed at a ridge axis have been partially buried by sediment,
as in area B. Much higher valuesare found in areasof extreme
rejected; that is, the significance of the correlation [Press et topographicvariation, such as at the continentalshelf breaks
al. , 1986]. and the trenches.
The scatterplots (bottom panels of Figures 6a, 6b, and 6c)
show the x value we determined for these areas. Also shown
Results
are solid and dashedlines. The solid lines are drawn through
theoriginwitha slope
givenby-I-Oh/Og,
Obeing
estimated
by The S(x) grid obtained above was combined with the
the Z method and the sign being taken from the sign of x. The downwardcontinued,band-passedgravity g(x) and low-passed
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,813

O* 30* 60* 90* 120* 150* 180* 21 O* 240* 270* 300* 330*
-30* -30*

-40'

-50*
ill ::5: "
-60*
.JJ
:5*:5:5:::::
..
::::: . ::::
... .
... :::
.::::
.:.:. .:.:. ... :.:.:
-70

Confidence
thatHoisfalse
0.00 0.01 0.05 0.10 0.50 0.90 0.95 0.99 1.00

::::'. :.' .....'.......... : ..............

?:'.:,.............
'",.... . . ......:.::'.: : ':-:
'.-:' "':.::............ '"-:';.;j'&..-.:.-
....'.....:-:*
...... '......
' :.... '-5::::a
::.

-40* J" ....'

.' .....
':'::'.:.;'...::.:
"::::'
"'"........
':'::"'' : .........":.:.::'..:::.::.:.
::"" "':':::'
..... ":5::*'g*:::' ':z' :: :5: " ......'" ::::::':
' ....::'.:.:::5:*
%.'....................
':"":::'
.........
"'*::' '.......... '"' ' ':: ........
:.:..
.
.....
:
:':.............
:...'"::::
'::::z:..':::.:.:.:."
.... :i$;:5
...........
:::::......'-::
'3:'-':
'.':..:."
......... '..........
.:::::..-:::..:::::
:::::
::..::::..
:...: .....
:.:.':5:.::
...
........
'.:-'-.::::
: ......
:5::'.::
,:
:::::.:.:
.....................
.......
:::.: :::z:.
:.::"'.".
:::::.:::."
......
..:................... .:.
.......... :::
.......
....................
============================
.".:.... ::'" "..
-40* ::: :..
.....:..'
:; ......
........ :..':i)
........
.... -'. ...... ........
...... .' '. ...... :.... '.'
....:',.' :. --'
" ....... '. :.:::'5:::::-.-')i
.;-'::
...... ..... .................
::-'-:-:-m*:::::::'""::"
':*.
...... ::.::
" ' .' .....
....',-::
'" . :>':-::.::'.
':....
: .:::"
:': ................

:':-".-::": -.,:.'-
.......
:: ' :..,,:::..,
':e: ........ ::::
....3 -.- :::: . ' .:.:.::'..:5,
.....................
'. --'.'...:.:'-::5.-
....... -.::::-::.:.:
.............
?::;::.:.:':':5:::?
.:.-'-'--:
-::......
....':.?
..:
................................
--:- .....
g...... :....
'.':-:.::'
:::5::..'.:
':-: ..........
':":. 5: ?':..
.--v. :::.:::.,'.:5
......: ..... '":': . -:.-.:::'::....... .:;
,....:.4
:'/ ...................
.?(?':-
I:?:'.-.,:':?:::V
-':' :'::?::::::::g:2(:'":;:?:
.....
---%d" %...,:.:.:.::
.:.:.:::?::-:'::;.'.:;"
,:::::::
.............................
........... ':::: .........
4..:-,'::
::'::'":"..--
....:.: ::::::
'": '-,:,:
:'":::'.: 'A/
'"'.-: ''::'.....
--:::,::-
--..........
'........
: , ')
..:-
':':':
::-,::::"::-:.::::-:::::,--';.-::':.:::::::::gq
:g::: .. :':.:.:::<:.:::
.......:':
...............
:'':
'::':::' ..................
.......
'...........
;':'":';::::;:/?Vi";::'
.......
::::.::'*'.-:::
:::..'""-.::'
:...........
-.- ...............
::'::::'::-':'".'
,':*:.';:;:.'"'-".,.;...
-50'
.........
:
.2[gl
. z -.: ;.y::::..:::,.
.-.::-:::-.:'..
:- ::'[::::[:.;
:::[:'-.-
':: :k:3
.........
:.?2:2'::.[$
.:":::':::'?-'2
:::::::::::::::::::::::::::::::
.:' ::' :-:' ::':.:.z'.' .:e4:::::
'::.:.':::'
:'"::::':':.':-
;x-- 'm-':..'-':.-
......
- ..:-'.'
I';....::,:.:-:::':: . ,. ........
' ............... =======================
..::
...............
":.,-,? ---::..
-:::......
".': :' ........
'-?::.':.' ..:
::
....... -sine'.--:-:-
............
... '1
,::,,
-60 ...... . ........ .... -::?:z://....:...;::
..,:.:.:::
.........
:.:::... ........ .......
......
.......
.. ::.::::.
.
.:......
'.::....
. .....
..... ::::' '-:
....
,:.>... : ....
..,....:.
.....
.:::.:..:.::.:::-
........
::;::
....... ..........
....:.:..:
.....m:.. ..;
.....
:.:.;,...:.::::::::.:...:.;.:::::..
......
: .............
::::. -......
.::-:?.:.::
:::...:.::.::-:::,.
::: -::.'..:,:: -. ............
........
:.-...-. . _60
..::.:::...:..:....
::: ....
..................
.....
..... .. ......
- ::-... ..: .:..:::.:..,....:/::'
:::....::-
.. ::::-..: ........:.,
:..:::.'.'., ...::: .:::..:::,--..::::.:.-.:....:::3
........:.......... .::::.:
.... :. :..::. :.......
v
........
:.::.... ,. .:.:.:..:......
::..: .-..:
.....:::::::::::::::::::::
........... ::-:::.:::..s::x:
:.:...:::::.2';m>:
.:: ....
::::.::::: ..:::
....... :--.:::::.
:::: ...../&
........:..:.:
.........
.....:::,...-:--:,..:..::....:
.:.-,......:::::::::
:.::,:
:.:)..--':'3,
: '::::'.'-::':':
'.':Y:. ..'.:-.'. :-' '.:.:'::..-.'3:>':".:'
.:::-, ; .$ :..4..... ::.e::::*.:'.::':':':.:;':
:'
.................
.:5 .....-:..:"
::::::::::::::::::::::.:...-'-:8:::-..:',.L'::I
.::::
'--5:-::-:.
-..:::
:.:??- ...............
z./ '...y. :, ::-."
.. : --"
, :5.:::
...:'"-.::
' .....':':'.....
. .'.'-::
.. .-:.
_70 =======================
;;:::::::::?:::-?.:'
:.... 57:....:. :..............
....:.:
.........
:;::-:-::.:-:
::.;..:.

Correlation Coefficient,
-1 .o o.o 0.2 0.4 0.6 0.8 1.0

-30* -30*

-40
..... "i"":i?'"'"f':''"""d:!i
'"'"::""':::'
!!i':*
'"'"
''
' '" -40'

_50 _50

-60
-70 .... '"'"":"...."4..".-"
-60 70

Slope,S (m/mGal)
0 6 13 16 23 30

' ''-.::''
'iii"."..-'-:...
:gf / -iF
.ii.."...ii.."...i!i?:
-40' . ,p 'ii:}i!
' AAP -40'
-50* ' -50*

-60' SlAP ,,, ... j-60*

0* 30* 60* 90* 120* 150* 180* 210 240' 270' 300' 330'

Figure 7. Results of "inverseNettleton" procedure(top three panels), and correspondinggeologic features

(bottompanel). AAP, BAP, EAP, SIAP, and WAP are the Argentine,Bellingshausen,Enderby,SouthIndian,
and Weddell Abyssal Plains. BR and LR are the Broken and Louisville Ridges. Lines indicate plate
boundaries.
21,814 SMITH AND SANDWELL: BATHYMETRIC
bathymetry d(x) to form the prediction by equation (1). The
results are shown as Mercator projected images in Plates 2
through 5. A large format polar stereographic projection
poster is available from the National GeophysicalData Center
and the digital grid file is available by anonymousftp from
Internet site ftp.ngdc.noaa.gov[Smith and Sandwell, 1994].
It is beyond the scope of this paper to present a detailed
comparison between the predicted bathymetry images and
previously published bathymetric maps such as the General
Bathymetric Chart of the Oceans (GEBCO) [Canadian
Hydrographic Service, 1981] or ETOPO-5. Moreover, there
are many detailed local bathymetric maps that should be
considered in such a comparison [e.g., Lonsdale, 1988;
Mammerickx, 1992]. Using the digital data grid file, the reader
can prepare a local map at the scale and projectionneededfor
such quantitative comparisons. There are, however, some
general features that should be noted, especially in relation to
the published gravity anomaly maps.
Most short- and intermediate-wavelength features (15-160
km) apparentin the gravity map [Sandwell and Smith, 1992]
are also apparent in the predicted bathymetry map. The main
differences are that the bathymetric prediction contains
considerablelong-wavelengthinformation and the downward
continuation procedure amplifies the shortest-wavelength
gravity structure with respect to the intermediate-wavelength
gravity structure. The combination of these has a profound
effect on the appearance of the bathymetric map. For
example, the gravity anomaly map shows large-scale struc-
tures such as continental margins and oceanic plateaus as a
gravity low seawardof the shelf, a gravity high inboard of the
shelf and a decay of the gravity toward zero as one moves
further inboard of the shelf. This is the classic gravitational
edge effect of a compensatedmargin. When the ship sound-
ings and gravity are processed and combined as described
above, these dipolar edge effects change to the true steplike
edge. There are many examplesof this throughoutthe map;
the most prominent is the Campbell Plateau (45-55Slati-
tude, 165-180 longitude;see also Plate 1). The gravity map
shows gravity lows on shallow parts (< 1000 m deep) of the
plateau, while the bathymetry map clearly reflects the high
plateau topography.
A second major change between gravity and bathymetry
occurs at isolated loads such as seamounts and plateaus
containing imbedded seamounts. On the gravity map, these
structures appear as an isolated gravity high ringed by a
prominent gravity low. As shown in Figure 4 the gravity low
reflects the componentof gravity causedby downwardflexure
of the Moho; most seamounts do not have prominent
bathymetricmoats. In the above analysiswe tried to avoid the
longer wavelengths associatedwith lithospheric flexure in
order to suppress these negative sidelobes. A visual
examinationof the bathymetricchart showsthat only a few of
the predicted seamounts have moats. Most show conical
shapesthat better reflect the true seafloor morphology. For
example, predicted seamounts on the Louisville Ridge
compare well with actual bathymetric profiles [Lonsdale,
1988].
There are at least two other classes of tectonic features that
are enhanced and clarified by the above processing. First,
while the detailed location of the spreading ridge axis is
apparent in almost all locations on the gravity anomaly map,
there are a couple of areas where the ridge axis location is
ambiguous. For example, in the gravity map there is no clear
PREDICTION FROM ALTIMETRY
ridge expressionof the southernmostMid-Atlantic Ridge (45 -
55S). The bathymetric map contains longer-wavelength
topographic information that creates a symmetrical crest
which can be used to infer the approximateridge axis location;
morphology considerations can be used to trace the axial
valley to 50S where it ends abruptly and becomesa broad
axial dome perhapsassociatedwith the Bouvet Hot Spot. A
similar situation occurs along the SoutheastIndian Ridge at
--80 longitude where, using symmetry arguments,an axial
valley can be located to the southeastof Amsterdam and St.
Paul Islands.
Second, sharp linear structures such as transform fault
valleys, transverseridges, and fracture zones become sharper
and more continuousafter the above processing. For example,
the long transverseridges of the Heezen, Tharp, and Udintsev
Fracture Zones (South Pacific) are readily apparent in the
bathymetricchart, while they were sometimesobscuredin the
gravity chart by longer-wavelength information perhaps
related to lithosphericflexure or thermal isostasy.
While in many cases the bathymetric map clarifies the
tectonic structures,it hides many important features that are
buried by sediments. Thus the gravity anomalymap may be
better for investigatingthe tectonicsof the old sedimentary
basins.
Validation
In January and February of 1992, the R/V Maurice Ewing
made a survey acrossthe Pacific-Antarctic Rise near Fracture
Zone XII (now also called the Pitman Fracture Zone) using
Global Positioning System navigation and Hydrosweep
multibeam sonardata. These data have not yet been releasedto
the National GeophysicalData Center and were not available
to us when we compiled the soundingswhich form our grids.
After we had made our prediction,we obtainedthe centerbeam
depths from the Hydrosweep survey (S.C. Cande, personal
communication, 1993) to compare with our prediction. These
data lie in an area of holes in our survey coverage (Figure 1)
where our prediction was poorly constrained,so they can put
the prediction to a strong test.
Figure 8 (top) shows a profile acrossthe ridge; the profile
location is shown in Plate 1. The small crossesare soundings
from the Ewing, and the curve is drawn showingour prediction
at the correspondingpoints. The Hydrosweep data are very
dense, and we have decimated them to only one point per
kilometer in Figure 8 for clarity. The bottompanel of Figure 8
showsthe Ewing soundingsand the ETOPO-5 data for compar-
ison. The prediction resolves many details not seen in
ETOPO-5, includingthe correctlocationof the ridge crest.
We sampledour predictiongrid at the Ewing data pointsfor
11,748 km of the cruise, and used a linear interpolant to
sample both data at 1 km intervals; the two series were then
fed to a cross-spectralanalysis program (spectrumld of the
Generic Mapping Tools (GMT) system of Wesseland Smith
[1991]). The program employs Welch's [1967] method to
estimatespectra;we used a window lengthof 1024 points. We
also formed the difference between the Ewing data and our
prediction and estimated its spectrum by the same method.
The resultsare shown in Figure 9. The power spectraldensi-
ties (top) are similar at wavelengthslongerthan 30 km or so,
while at shorter wavelengths the prediction has much less
power than the Ewing data. This is to be expectedfrom the
high-cut propertiesof W(k) (middle). The gain (middle) and
SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,815
21,816 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY
SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,817

II Ill
21,818 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY
 SMITH AND SANDWELL:BATHYMETRICPREDICTIONFROMALTIMETRY 21,819

Distancefrom port (km)

2400 2600 2800 3000 3200 3400 3600 3800

-2

+
++

+ ' ++
+ + +
-3 +
+

++,

-4

Ewing 92-01 and Prediction

-5

, I , , , I , , , I , ,

4.+

+ +

/ I;+
+ .++ +

' ' ' i ' ' ' i ' ' ' i ' ' I ' ' ' i ' ' i ' ' i ' '

2400 2600 2800 3000 3200 3400 3600 3800

Distancefrom port (km)
Figure8. (top)Observed
shipsoundings(crosses)
fromR/V Ewing,whichwerenotusedto constrain the
prediction,
andtheprediction
(solidcurve).Location
of profile
is shown
in Plate1. (bottom)
Sameastop,
but solidcurveindicatescorresponding
ETOPO-5data,notprediction.

coherency(bottom)are calculatedassumingthat the Ewing tionis basedupond(x), whichhasbeeninterpolated

through
dataare the inputandthe predictionis the outputof a linear the holesin shipcoverage.Figure9 suggests thatthe interpo-
system,with noisein the output.The gainandcoherency are latedsolutionis not very goodat thesewavelengths.
low at wavelengths shorterthan30 km, againbecauseof the
highcutin W(k). An interesting featureof boththe gainand Evaluation and Discussion
coherency plotsis thattheseare slightlyhigherin the 100-30
km rangethantheyarein the 250-100km range. The former We computedthe misfit betweenour predictionand our
is withinthe passbandof W(k), wherethe prediction relieson griddedsoundingsat all grid pointsconstrained
by data. A
S(x) andg(x), while the latteris the rangewherethe predic- histogram
of thedifferences is shownin Figure10 (top). The
21,820 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

15000 I I I I
Wavelength(km)
103 102 10
10s i I .... I I I I [111 i i I i

107 -
10000

106 -

105 -
5000

104 - Observed

103 - Predicted
0
- 1500 - 1000 -500 0 500 1000 1500
102
Difference (m)

1.0 / , I , I I I I
1.0

0.5

! o.t ,
',_
0.4
/-]
.......
i.......
i-..rnedian=-2.6rn
'"i'"
....
......
...... --.i...
0.2
0.0

1.0 0.0
0 200 400 600 800 1000

AbsoluteDifference(m)
Figure 10. (top) Histogramof differencesbetweenprediction
and ship soundingsshows a distributionmore peaked and
longer tailed than the normal distribution. (bottom)
Cumulative distribution of absolute differences shows that the
difference is less than 100 m 50% of the time and less than
240 m 80% of the time.

distribution appears to be symmetric around zero with a

sharperpeak and longer tails than a normal distribution. The
mean difference is 7.8 m and the median difference is -2.6 m.
The root-mean-squarevalue (rms) is 257 m and the median
0.0
absolutevalue (mav) is 96 m. As explainedabove, if the
103 102 10 distribution were normal, we would expect 1.48 times the
median absolutevalue to equal the root-mean-square value
Figure9. Cross-spectral analysisof Ewing dataandpredic- (assumingthe distributionto be locatedat zero); clearly, the
tion. (top) Power spectraldensity. (middle) Gain (coherent rms is much larger due to the long-taileddistribution.
amplitudeof predictionrelativeto Ewing data),with W(k) for Becauseof theselong tails, the rms and mav do not fully
d=5.5km, the averagedepthof thedata. (bottom)Coherency describethe distribution,and so we alsocomputedthe cumula-
betweenpredictionandEwing data. tive distribution of the absolute value of the differences
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,821

(Figure 10, bottom). With this we can seethat more than 80% Here the ridge and trough of a major fracture zone meets a
of the differences are less than 257 m, the rms value, whereas a
thickly sedimentedarea; sedimenthas nearly filled the trough,
normal distributionwould have only 68% at this point. Since while the ridge is still exposed. This has the effect of clipping
the median absolute value is 96 m, 50% of the differences are the negativeh valueswhile making little changein positiveh
lessthan or equal to this value. values, leading to the skewed distribution in Figure 6b.
A map of the absolute differences is shown in Figure 11 Similarly, the positive gravity anomaliesare large, while the
(comparewith Figure 7). Shownhere are the averageabsolute negative ones are smaller. If the trough is filled completely,
differencesin 2.5 by 1.25 (135 km) "squares"as usedfor the the negative gravity anomalies will not vanish, becausethe
Nettleton grid. Squareswithout data are shown as having an sediment-basement density contrast will leave a small
averagedifferenceof zero. The very large differencesoccurin negative anomaly. This means that the gravity anomaly
areas of very rugged topography, notably the Scotia Sea, distributionmust always be less skewedthan the topography.
SouthwestIndian Ridge, Louisville and Broken Ridges, and The methodused here, and indeed any linear filtering method,
Hjort and Chile trenches. There are two reasonswhy this is makesno provisionfor a S value which dependson the sign of
expected. The first is that our 2.75-km grid has a Nyquist the gravity anomaly; it cannot predict topography with a
wavelength of 5.5 km, while W has a cutoff at 15-25 km, distributionmore skewed than the original gravity anomalies.
dependingon water depth, so there can be wavelengthsin the In the caseof area B, one may chooseS to fit the ridge and thus
observeddata grid which we will not predict. In areas of predict a too deep trough, chooseS to fit the trough and thus
rugged topography,the short wavelengthswhich are missing underpredictthe height of the ridge, or choosean intermediate
from the predictionwill be neededto fit the data. The second value and fit both of them badly. One of the reasonswe chose
reasonis that in someof theseareasthe local topographyh is slopesbased on the Z method was that it did a better job in
not small with respect to d, the regional variation. This is situations like area B than other methods we tried.
particularly true of features which rise to shallow levels. We are not sure what causesthe very large values of S we
Under these conditions the approximationthat g and h are observe in some areas. These seem to indicate large topo-
relatedby a linear filter is a poor one [Parker, 1973]. graphic featuresassociatedwith unusuallylow gravity anoma-
The theory given in this paper is based upon the lies. One explanationwould be that thesefeaturesare isostati-
assumption that gravity anomalies (in a limited band, cally compensated.We designedthe low cut of W (the stop-
perhaps)are due entirely to seafloortopography. If this is bandof Wl) to removewavelengths affectedby compensation,
true, then the slope parameter $ should be inversely assuming that all areas of the map would have a flexural
proportionalto p, the density contrastat the seafloor. Only compensation parameter3greaterthanor equalto 135 km. It
part of the range of $ values we found (Figure 7) can be is possiblethat some featureshave a more local compensation
sensibly inverted for p, however. $ in the range 13-16 (smaller30. Onecouldtry to increase thestopband of W1 to
m/mGal may be taken to indicate topographycomposedof guard againstthis, but then the passbandof W would become
basaltic rocks, and slightly higher values of $ may be very narrow, leading to other problems. We think our choice
appropriate for continental shelf and slope materials. of W is a good one. Most of the very large S values occur at
However, the very small, and perhaps also the very large, the edgesof plateausand continentalmargins,where there are
values of $ indicate conditions that the theory given here is abrupt and large changesin depth. When these changesare
not equippedto handle. band-passfiltered, the resulting h is probably a significant
The areasof low (< 10 m/mGal) $ and particularly of low fraction of d (for a Heaviside step function they would be
correlationandlow confidencein rejectionof Ho correspond nearly equal), so that the linear approximationof the gravity-
well with the areasof thick (> 200 m) sedimentin the isopach topographyrelationshipdoes not hold. However, the areasof
map of Hayes and LaBrecque [1991]. In these areas,the high S (Figure 7) are not well correlatedwith the areasof high
gravity anomaliesarise from subseafloorstructuresand not misfit (Figure 11), so nonlineareffects do not seem to entirely
from seafloortopography. At any point where thereis some accountfor the discrepancies.
sediment cover over other than flat basement,there may be The horizontal resolution of the prediction is also of
somecontributionto the gravity field from buried structures. interest. We want to know what is the smallest feature that can
Area B (Figure 6b) exhibits a particularlychallengingcase. be resolvedand how accuratelyparticularfeaturesare located.
oo 30 60 90 120 150 180 210 240 270 300 330

-30 ::::5.-'-':..;'-:-'5'i5::i
f'':-
-'-":'-'--'!
':- : .... 'il./''
'" ':'' ..... '""-'- m.-.,..
,i.
........
-30
" i, '" .,.::---:g
_40
o :
..............
:::.
:::::._:::"
:::::
....
:.:: .):&..::.'::..::
.::'
-.
::::(gg:::::::.:}::5::::}:
s...:::::::::::::::::::::::::::::::::::::::
40

_50 _50

.............................
-6 . ....
:] -60
_70 g;..;. -70

0 50 100 150 200 1000

Figure11. Averageabsolute
difference
betweenprediction
andshipsoundings in cellssameasin Figure7.
Manyareasshowdifferences
under50 m. Largedifferences
occurat areasof largevariations
in topography.
21,822 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

Because the wavelengthsshorter than the passbandof W are fore the variance of our prediction is less than that of the
missing from the prediction, this limits the resolution of the actual seafloor topography.
prediction. It is clear from Plates 1-5 that the resolutionis at We expectthat our predictionwill be usefulin plate recon-
least sufficient to distinguishaxial highs from axial valleys atstructions, analysis of regional trends in topography, and
mid-oceanridges. While this paper was in review, P. Lonsdale planningof oceanographic cruises. It shouldnot be used in
(personal communication, 1994) made a GPS navigated studies of the spectrum of topographydata, in particular
SeaBeam-2000 survey of the Eltanin and Udintsev Fracture roughnessand topography-gravityadmittancestudies. It can-
Zones, and found that predictedbathymetricfeaturesappeared not resolvepeak amplitudeswell, so the predicteddepthto the
within the width of the SeaBeam swath. Since the swath width summit of a seamountmay not be reliable; safety decisions
is about 3.4 times the water depth, this experiencesuggests about hazards to navigation should not be made from it.
that the location precision is better than _ 1.7d, or 8 km in However,it may be usefulin checkingthe reliability of appar-
deep water. The definitions of resolving power based on ently dubioussoundingson old charts.
length scalesin w(r) given above yield more conservative
(larger) limits. Appendix: Optimal Downward Continuation
As shown by equation (1) and discussedunder the inverse
of Gravity Data
Nettleton procedure,our methodbasesthe predictionon grav-
ity data in the passbandof W, sometimes(where significant The exponentialgrowth of downwardcontinuationrequires
negative correlations exist) in contradiction of the depth a high-cut filter for stabilization. The choiceof cutoff-wave-
soundingdata. Smith [1993] has shown that the majority of length is important;too much smoothingwill blur the details
data may be poorly navigatedin much of the southernoceans. of seafloor structure,while too little will swamp the prediction
In these situations,we believe the gravity data are more likely with noise. The choice can be formalized using the Wiener
to be correct than the ship data, and we make no attemptto fit theory [Wiener, 1949; Gonzalezand Wintz, 1987] to designa
the ship data. Thus our prediction cannot be taken as a filter which is optimalin the mean-square sense,and the cutoff
summary of, or substitutefor, original soundings. Scientists wavelength can be taken from the results of repeat track
who fit physicalmodelsto bathymetricdata shouldbe aware of coherency studies.
these facts. Since we have produced a prediction which is Supposethat we estimateg(x), the gravity field at a depthd,
perfectly correlated with gravity in the passbandof W wher- from our gravityobservations
at sealevel, go(X),usingthe
ever S is nonzero, any attemptsto interpret the gravity admit- downward continuation
operatorexp [2nkd]anda filter W2(k):
tance of our predictionwould be circular. Properusesof our
predictionmay include (but are not limited to) proposalof new O(k) = Go(k}exp[2;td] W2(k}. (A1)
surveys,reconnaissanceof tectonic settings,plate reconstruc-
tions, ocean current modeling, etc. Our prediction cannot The circumflex on G in (A1) indicatesthat this is an estimate
contain very short wavelengths and underestimatestopogra- of the downwardcontinuedgravity field. Supposethe true field
phy in ruggedareas(Figure 11). We do not guaranteethat the at depthd is given by
actual topographyof the oceanfloor is as we have predicted.
G(k) = S(k)exp2kd, (A2)
Conclusions
where s(x) is the "signal"and our observedgravity contains
Flexural isostaticcompensationtheory suggeststhat down- both signal and noise;
ward continued gravity anomalies should be correlated with
seafloor topography in approximately the 15-160 km wave- go (x) = s(x ) + n(x ). (A3)
length band. The theory representsa linear approximationto
a nonlinear problem, and the underlying assumptionof spatial
(In this appendixwe use S for signal and N for noise; else-
invariance must be relaxed in order to accommodateregional
where S is the slope from the inverse Nettleton procedure.)
changes in geology. Stable, high-resolution inversion of
Now the integratedsquarederrorin our estimateE is
gravity data for bathymetricpredictionrequirescareful design
of filters to suppressnoise in downwardcontinuation.
We find significant correlations between downward
continuedgravity and topographyafter thesehave been band-
pass filtered, in areas of rough seafloor topographyor where
sedimentcover is less than 200 m thick. In more heavily sedi-
(the secondequality in (A4) follows from Rayleigh'stheorem
mented areas, the seafloor appears quite flat, while gravity
variations persist; the latter may representbasementtopogra- [Bracewell, 1978]), andthe Wienerfilter is that W2 which
minimizesE. When we substitutethe first threeequationsinto
phy buried under the sediments. In areasof partial sediment
the fourth, there are signal-noisecrosstermsSN* and NS* (the
cover, where basementhighsare exposed,while lows are filled
asterisk indicates a complex conjugate); these are Fourier
in, the relationshipbetweengravity and bathymetrycannotbe
transforms of the cross covariance between the signal and
adequatelyexpressedby a linear operationasusedhere.
noise. Assuming that these are uncorrelated, these terms
We have predictedbathymetryin previouslypoorly charted
becomezero. We assumealso that the power spectraof the
waters. Our predictionis validatedin an area of the Pacific-
signal and the noise are isotropic and thus functionsof k, not
Antarctic Ridge by a GPS-navigatedHydrosweepsurvey. The
k, so that
prediction resolvesthe location of important tectonic features
suchas ridge axes better than previouslyexistinggriddeddata
productssuchas ETOPO-5. The predictionis band-limitedand E=2t exp
[4tkd]
{[S(k)[[1-W(k)]-N(k)[W(k)}dk.
(A5)
cannotresolve sharpfeatureswith their full amplitude;there-
 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY 21,823

A stationary
pointof E withrespect
to W2 is foundby setting Is()12= exp[-4nkd], (All)
aE/aW2 = O:
whereA 1 is someconstant.Let the noisebe white (constant
in wavenumber) in the altimeter sea surface height measure-
0=i)E
_4Iexp
w Jo
[4kd]
{IN(k]2W2(k)qS
(k]2[
1-W2(k)]
}dk.
(A6) ment, so that it is proportional to k in the derivative gravity
data; then
This point is a minimumfor E, which is shownby differentiat-
ing again: IN(k)12=A2k2, (A12)

)2
E whereA2 is anotherconstant.Undertheseassumptions,
0<
=4nIexp
E4ckd
{IN()I
2+Is()I
2 R{k)= A k4exp[4nkd, (A13)

Now by the calculus of variations, (A6) is satisfied when the
integrandis zero, or
+ (AS)
where
Je IN()12 (A9)
Is()l
is the amplitude-squarednoise-to-signalratio as a function of
wavenumber.Note that W2 doesnot dependon the factor
exp[4nkd]; the Wiener filter is independentof the convolution
operatorwhichrelatesg andgo'
Sandwell and McAdoo [ 1990] have examinedthe coherency
between pairs of Geosat altimeter time seriesalong repeating
groundtracks of the Exact Repeat Mission. The coherencyin
these data can be used to estimate R(k) for the Geosat altime-
ter. Since the Sandwell and Smith [1992] gravity field is
dominatedby the Geosat Geodetic Mission data, we expect R
for the gridded gravity field to be similar to that obtainedby
Sandwell and McAdoo [1990]. If we assume that the two
where A is an unknown constant. If we choose A = 80,000
km4, thenwhenwe substitute
(A13)into(A10)we get= 0.5
when d = 4 km and k = 1/(30 km), the predictedform for (k)
matchesFigure 6b of Sandwell and McAdoo [1990], and the
wavelengthsat which (k) = 0.5 vary with water depth in the
manner describedby them. (Small increasesin R in areasof
large currentswere also detectedby Sandwell and McAdoo and
are not describedby (A13).) The Sandwell and Smith [1992]
gravity field showscoherencywith ship gravity measurements
[Smith et al., 1993; Neumann et al., 1993] at shorter wave-
lengths than the Sandwell and McAdoo study would suggest,
and it appearsthat in shallow water over ridge axes the field
has short-wavelengthdetailswe want to preservein our predic-
tion;we thereforechoseA=9500 km4 ratherthan80,000km4.
Note that the derivation leading to (A8) shows that the
Wiener filter is independent of the downward continuation
term; however, because R(k) depends on depth via (A13),
W2(k)implicitlydepends
on d.
Acknowledgements. R. L. Parker offered many helpful comments
throughoutthis research. B. T. May raisedthe questionof resolving
powerand suggested the opticalRayleighcriterion. Many calculations
and illustrations were made with the GMT software of Wessel and Smith

samples of altimeter time series have the same signal and [1991]. The manuscript was improved with suggestionsfrom A.
Goodwillie, K. M. Marks, and D.C. McAdoo and reviews by G. A.
noises uncorrelatedwith the signal and with each other, and
Neumann and P. Wessel. This work was supportedby the National
the two noises have the same spectrum, then the spectral
Science Foundation, Division of Ocean Sciences(OCE 92-17164).
coherencyis

(k)=[1+ R(k)-2 (A10)
[Bendatand Piersol, 1986]. The coherencyis near 1 when R is
small and near 0 whenR is large;Sandwelland McAdoo [ 1990]
gave plots of and estimatesof the wavelength(l/k) at which
= 0.5. They found that is near 1 for long wavelengthsand
near 0 for short wavelengths,with the transitional wavelength
in the range 20-30 km. They showedthat the wavelengthfor
= 0.5 varies with the depth to the oceanfloor; in deep water
(d = 4 km) this wavelengthis about 30 km, while in shallower
water (d = 2.5 km) this wavelength is about 20 km. (The
values given here are worked out from their results;they give
only one example of coherencybetween two individual time
series, and the remaining examples are coherency between
pairs of yearly average time series. The wavelengths for =
0.5 given here are appropriatefor use in equation(A10).)
Here we create an analytic expressionfor R(k) which when
used in (A10) yields (k) consistent with the results of
Sandwell and McAdoo [1990]. Our expressionis inspired by
the following guessesabout S(k)and N(k): Let the gravity
signal arise from applicationof (5) to a quasi-fractal,random
walk topography
of the formH(k) proportional
to k-1 [see
Malinverno, 1989, and references therein]; then
References
Backus, G., and F. Gilbert, The resolvingpower of grossEarth data,
Geophys.J. R. Astron.Soc., 16, 169-205, 1968.
Banks,R. J., R. L. Parker,and S. P. Huestis,Isostaticcompensation
on a
continentalscale:Local versusregionalmechanisms, Geophys.J. R.
Astron. Soc., 51,431-452, 1977.
Baudry,N., M. Diament,and Y. Albouy,Preciselocationof unsurveyed
seamounts in the Austral archipelago area using Seasat data,
Geophys.J. R. Astron. Soc.,89, 869-888, 1987.
Bendat, J. S., and A. G. Piersol, Random Data Analysis and
MeasurementProcedures,2nd ed., 566 pp., JohnWiley, New York,
1986.
Bracewell,R., The Fourier Transformand Its Applications,2nd ed., 444
pp., McGraw-Hill, New York, 1978.
Brownlee, K. A., Statistical Theory and Methodology in Scienceand
Engineering,2nd ed., 590 pp., JohnWiley, New York, 1965.
Caldwell, J. G., and D. L. Turcotte,Dependenceof the thicknessof the
elasticoceaniclithosphereon age, J. Geophys.Res.,84, 7572-7576,
1979.
Canadian Hydrographic Service, General Bathymetric Chart of the
Oceans (GEBCO), 5th ed., HydrographicChart DistributionOffice,
Ottawa, Canada, 1981.
Carr, M. H., G. Greeley, K. R. Blasius,J. E. Guest,and J. B. Murray,
Some Martian features as viewed from the Viking orbiter, J.
Geophys.Res.,82, 3985-4015, 1977.
21,824 SMITH AND SANDWELL: BATHYMETRIC PREDICTION FROM ALTIMETRY

Cochran, J. R., An analysisof isostasyin the world's oceans,2, Mid- NumericalRecipes,818 pp., CambridgeUniversityPress,New York,
oceanridge crests,J. Geophys.Res.,84, 4713-4729, 1979. 1986.
Constable, C., and R. Parker, Deconvolution of long-core Rousseeuw,P. J., and A.M. Leroy, Robust Regressionand Outlier
palaeomagnetic measurements-Spline therapy for the linear Detection,329 pp., JohnWiley, New York, 1987.
problem,Geophys.J. lnt., 104, 453-468, 1991. Sailor, R. V., and M. L. Driscoll, Noise models for satellite altimeter
Craig, C. H., and D. T. Sandwell,Global distributionof seamounts
from data,Eos Trans.AGU, 74(16), SpringMeetingsupp!., 99, 1993.
Seasatprofiles,J. Geophys.Res., 93, 10,408-10,420, 1988. Sandwell, D. T., A detailed view of the South Pacific geoid from
Dixon, T. H., M. Naraghi,M. K. McNutt, and S. M. Smith,Bathymetric satellitealtimetry,J. Geophys.Res.,89, 1089-1104, 1984.
predictionfrom Seasataltimeter data, J. Geophys.Res.,88, 1563- Sandwell, D. T., Antarctic marine gravity field from high-density
1571, 1983. satellitealtimetry, Geophys.J. lnt., 109, 437-448, 1992.
Dorman, L. M., and B. T. R. Lewis, Experimentalisostasy,1, Theory of Sandwell,D. T., and D.C. McAdoo, High-accuracy,high-resolution
the determinationof the Earth'sisostaticresponseto a concentrated gravityprofilesfrom 2 yearsof the Geosatexactrepeatmission,J.
load, J. Geophys.Res., 75, 3357-3365, 1970. Geophys.Res., 95, 3049-3060, 1990.
Ford, P. G., and G. H. Pettengill,Venus topographyand kilometer-scale Sandwell, D. T., and G. Schubert, Geoid height-age relation from
slopes,J. Geophys.Res.,97, 13,103-13,114, 1992. Seasat altimeter profiles acrossthe Mendocino Fracture Zone, J.
Goodman,J. W., Introductionto Fourier Optics,287 pp., McGraw-Hill, Geophys.Res., 87, 3949-3958, 1982.
New York, 1968. Sandwell,D. T., andW. H. F. Smith,Globalmarinegravityfrom ERS 1,
Goodwillie, A.M., and A. B. Watts, An altimetricand bathymetricstudy Geosat, and Seasat reveals new tectonic fabric, Eos Trans. AGU,
of elastic thickness in the central Pacific Ocean, Earth Planet. Sci. 73(43), Fall Meeting suppl.,133, 1992.
Lett., 118, 311-326, 1993. Siemens,C. W., On determiningthe depthof the seawithoutthe useof a
Gonzalez,R. C., and P. Wintz, Digital Image Processing,2nd ed., 503 soundingline, Philos. Trans.R. Soc.London,166, 671-692, 1876.
pp., Addison-Wesley,Reading,Mass, 1987. Smith, W. H. F., On the accuracy of digital bathymetric data, J.
Hayes, D. E., and J. L. LaBrecque, Sediment isopachs: Circum- Geophys.Res., 98, 9591-9603, 1993.
Antarcticto 30S,in Marine Geologicaland GeophysicalAtlas of the Smith, W. H. F., and D. T. Sandwell, Charting the remote southern
Circum-Antarctic to 30S,Antarct. Res. Ser., vol. 54, editedby D. E. oceans with satellite altimetry, Eos Trans. AGU, 73(14), Spring
Hayes, AGU, Washington,D.C., 1991. Meeting supp!., 1992.
Jung,W.-Y., and P. R. Vogt, Predictingbathymetryfrom Geosat-ERM Smith, W. H. F., and D. T. Sandwell, Sea floor topographypredicted
and shipborneprofilesin the SouthAtlantic Ocean,Tectonophysics, from satellite altimetry and ship depthmeasurements, Rep. MGG-09,
210, 235-253, 1992. World Data Cent. A for Mar. Geol. and Geophys.,Natl. Geophys.
Lewis, B. T. R., and L. M. Dorman, Experimental isostasy,2, An Data Cent.,U.S. Dep. Commer.,Boulder,Colo., 1994.
isostaticmodel for the U.S.A. derivedfrom gravity and topographic Smith, W. H. F., and P. Wessel, Gridding with continuouscurvature
data,J. Geophys.Res., 75, 3367-3386, 1970. splinesin tension,Geophysics, 55, 293-305, 1990.
Lonsdale,P., Geographyand history of the Louisville hotspotchain in Smith, W. H. F., H. Staudigel,A. B. Watts, and M. S. Pringle, The
the southwestPacific, J. Geophys.Res., 93, 3078-3104, 1988. Magellan Seamounts: Early Cretaceousrecord of the South Pacific
Malinverno, A., Testing linear models of sea-floortopography,Pure isotopicand thermal anomaly,J. Geophys.Res., 94, 10,501-10,523,
Appl. Geophys.,131,139-155, 1989. 1989.
Mammerickx, J., The Foundation Seamounts: Tectonic setting of a Smith, W. H. F., D. T. Sandwell, K. M. Marks, and D.C. McAdoo, On
newly discoveredseamountchainin the SouthPacific,Earth Planet. the accuracy of marine gravity fields calculated from satellite
Sci. Lett., 113, 293-306, 1992. altimetry,Eos Trans.AGU, 74(16), SpringMeeting supp!.,99, 1993.
Marks, K. M., and R. V. Sailor, Comparisonof GEOS 3 and Seasat Van Wyckhouse, R., SYNBAPS, Tech. Rep. TR-233, U.S. Natl.
altimeter resolutioncapabilities,Geophys.Res. Lett., 13, 697-700, Oceanogr.Off., StennisSpaceCenter,Miss., 1973.
1986. Watts, A. B., An analysisof isostasyin the world'soceans,1, Hawaiian-
Marks, K. M., D.C. McAdoo, and W. H. F. Smith, Mapping the Emperorseamountchain,J. Geophys.Res.,83, 5989-6004, 1978.
Southwest Indian Ridge with Geosat, Eos Trans. AGU, 74, 81, 86, Watts, A. B., On geoidheightsderivedfrom Geos3 altimeterdataand
1993. flexure of the lithospherealong the Hawaiian-Emperor seamount
McKenzie, D. P., and C. Bowin, The relationshipbetweenbathymetry chain,J. Geophys.Res.,84, 3817-3826, 1979.
and gravity in the Atlantic Ocean,J. Geophys.Res.,81, 1903-1915, Watts, A. B., The strengthof the Earth'scrust,Mar. Technol.Soc.J., 17,
1976. 5-17, 1983.
McNutt, M. K., Compensationof oceanictopography:An applicationof Watts, A. B., J. H. Bodine, and N.M. Ribe, Observationsof flexure and
the responsefunction techniqueto the Surveyor area,J. Geophys. the geologicalevolutionof the Pacific Oceanbasin,Nature, 283,
Res., 84, 7589-7598, 1979. 532-537, 1980.
National GeophysicalData Center, ETOPO-5 bathymetry/topography Welch, P. D., The use of the fast Fourier transformfor the estimationof
data, Data Announc. 88-MGG-02, Natl. Oceanic and Atmos. power spectra: A method based on time averagingover short,
Admin.,U.S. Dep. Commer.,Boulder,Colo., 1988. modified periodograms,IEEE Trans. Audio Electroacoust.,A U-15,
National Geophysical Data Center, GEODAS CD-ROM worldwide 70-73, 1967.
marine geophysicaldata, updated2nd edition, Data Announc. 93- Wessel, P., and W. H. F. Smith, Free softwarehelps map and display
MGG-04, Natl. Oceanic and Atmos. Admin., U.S. Dep. Commer., data, Eos Trans. AGU, 72, 441,445-446, 1991.
Boulder, Colo., 1993. White, J. V., R. V. Sailor, A. R. Lazarewicz, and A. R. LeSchack,
Nettleton, L. L., Determinationof density for reductionof gravimeter Detection of seamountsignature in Seasat altimeter data using
observations,Geophysics,4, 176-183, 1939. matchedfilters,J. Geophys.Res.,88, 1541-1551, 1983.
Neumann, G. A., D. W. Forsyth, and D. Sandwell, Comparison of Wiener, N., Extrapolation,Interpolation,and Smoothingof Stationary
marine gravity from shipboardand high-density satellite altimetry TimeSeriesWith EngineeringApplications,JohnWiley, New York,
along the Mid-Atlantic Ridge 30.5-35.5S,Geophys.Res. Lett., 20, 1949.
1639-1642, 1993.
Parker, R. L., The rapid calculationof potentialanomalies,Geophys.J. D. T. Sandwell,ScrippsInstitution
of Oceanography,
Code0225, La
R. Astron. Soc., 31,447-455, 1973. Jolla, CA 92093-0225.
ParsonsB., and J. G. Sclater,An analysisof the variationof oceanfloor W. H. F. Smith, NOAA, Code N/OES-12, SSMC-4, Station8423,
bathymetryand heat flow with age, J. Geophys.Res.,82, 803-827, 1305 East-WestHighway,Silver Spring,MD 20910.
1977. (ReceivedDecember22, 1993; revisedApril 11, 1994;
Press,W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, acceptedApril 12, 1994.)