Slip distribution of the 2003 Tokachi‐oki Mw 8.1 earthquakefrom joint inversion of tsunami waveforms and geodetic data

F. Romano,1 A. Piatanesi,1 S. Lorito,1 and K. Hirata2

Received 4 June 2009; revised 23 July 2010; accepted 24 August 2010; published 23 November 2010.

[1] We study the 2003 Mw 8.1 Tokachi‐oki earthquake, a great interplate event thatoccurred along the southwestern Kuril Trench and generated a significant tsunami. Todetermine the earthquake slip distribution, we perform the first joint inversion of tsunamiwaveforms measured by tide gauges and of coseismic displacement measured both byGPS stations and three ocean bottom pressure gauges (PG) for this event. The resolutionof the different data sets on the slip distribution is assessed by means of severalcheckerboard tests. Results show that tsunami data constrain the slip distribution offshore,whereas GPS data constrain the slip distribution in the onshore zone. The three PG dataonly coarsely constrain the offshore slip, indicating that denser networks should beinstalled close to subduction zones. Combining the three data sets significantly improvesthe inversion results. Joint inversion of the 2003 Tokachi‐oki earthquake data leads tomaximum slip values (∼6 m) confined at depths greater than ∼25 km, between 30 and80 km northwest of the hypocenter, with a patch of slip (3 m) in the deepest part of thesource (∼50 km depth). Slip values are very low (≤1 m) updip from the hypocenter.Furthermore, the rupture does not extend on the plate interface off Akkeshi. As asignificant back slip amount (∼4 m) has accumulated there since the last 1952 earthquake,this segment could rupture during the next large interplate event along the Kuril Trench.

Citation: Romano, F., A. Piatanesi, S. Lorito, and K. Hirata (2010), Slip distribution of the 2003 Tokachi‐okiMw 8.1 earthquakefrom joint inversion of tsunami waveforms and geodetic data, J. Geophys. Res., 115, B11313, doi:10.1029/2009JB006665.

1. Introduction

[2] The Mw 8.1 Tokachi‐oki earthquake of 25 September2003 (Figure 1) has been one of the strongest and betterrecorded seismic events in the last 50 years in Japan, and thefirst large interplate earthquake ever recorded by the Japanesestrong motion networks K‐NET and KiK‐net [Kinoshita,1998; Aoi et al., 2000]. Large earthquakes occur frequentlyoff Hokkaido Island (northern Japan) because the PacificPlate subducts at the Kuril Trench just under the Hokkaidoregion (Figure 1). Several tsunamigenic earthquakes occurredin the last 50 years in this region, as for example the 1952Tokachi‐oki (M8.2), the 1958 Etorofu (M8.1), the 1963 offUrup (M8.1), the 1973 Nemuro‐oki (M7.7), and the 1994Hokkaido Toho‐oki (M8.1) [Piatanesi et al., 1999; Utsu,1999; Watanabe et al., 2006].[3] Many GPS stations recorded the on land surface dis-

placement due to the 2003 Tokachi‐oki earthquake. Thevertical component of the coseismic displacement was clearlycaptured also by three pressure gauges (PG) (two oceanbottom pressure gauges and one cable‐end station) installedon the seafloor near the Kuril Trench (Figure 1) [Mikada

et al., 2006]. Furthermore, the event generated a large tsu-nami along the southern coast of Hokkaido, with runupheights higher than 4 m at some locations (Figure 1) [Taniokaet al., 2004a]. The tsunami was recorded along the coasts ofboth Hokkaido and Honshu islands by the Japanese tidegauge network (Figure 1).[4] In previous studies, the 2003 Tokachi‐oki earthquake

has been analyzed following different approaches and usingvarious geophysical data sets. The seismic source, for example,has been investigated inverting tsunami traveltimes [Hirataet al., 2004], tsunami waveforms [Tanioka et al., 2004b],teleseismic data [Yamanaka and Kikuchi, 2003; Horikawa,2004; Robinson and Cheung, 2010], strong motion data[Honda et al., 2004,Aoi et al., 2008;Nozu and Irikura, 2008],GPS data [Miyazaki and Larson, 2008], teleseismic andstrong motion data jointly [Yagi, 2004], and finally combin-ing GPS and strong motion data [Koketsu et al., 2004].[5] In this study we perform, for the first time, as regards

this event, a joint inversion of geodetic and tsunami data. Ingeneral, fault planes of large subduction zone earthquakes liebeneath both land and ocean seafloor. Thus, tsunami wave-forms and PG data can help constraining the slip distributionin the offshore zone, whereas geodetic data can mainly con-strain the slip onto the onshore zone [Satake, 1993]. There-fore, inverting them jointly might lead to the retrieval of abetter sourcemodel with respect to that estimated by invertingdata separately.[6] A description of the data used, their preprocessing, the

fault parameterization, and the methods for Green’s functions

1Department of Seismology and Tectonophysics, Istituto Nazionale diGeofisica e Vulcanologia, Rome, Italy.

2Seismology and Volcanology Research Department, MeteorologicalResearch Institute, Japan Meteorological Agency, Tsukuba, Japan.

Copyright 2010 by the American Geophysical Union.0148‐0227/10/2009JB006665

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 115, B11313, doi:10.1029/2009JB006665, 2010

B11313 1 of 12

calculation are provided first (sections 2, 3, and 4). Then, wedescribe the synthetic checkerboard tests used to assess dataresolving power for the slip distribution, both for single dataset inversions and for their joint inversion (section 5). Finally,we discuss the earthquake source characteristics of the2003 Tokachi‐oki earthquake retrieved from the inversion(sections 6 and 7).

2. Data

[7] Tsunami waves from the 2003 Tokachi‐oki earth-quake were recorded at Japanese coastal tide gauges. These

recordings are provided by Japan Meteorological Agency(JMA), the Hokkaido Regional Development Bureau (HRDB)of the Ministry of Land, Infrastructure and Transport, and theHydrographic and Oceanographic Department (HOD) of theJapan Coast Guard [Hirata et al., 2004]. In this study we use16 tsunami waveforms (Table S1 in auxiliary material andFigure 1).1 Particular attention is dedicated in making theazimuthal coverage around the earthquake source as large as

Figure 1. Location map of the 2003 Tokachi‐oki earthquake. Epicenter (white star) and focal mechanismare from Japan Meteorological Agency (JMA). Thin black lines are the surface projection of the subfaultsused in this study. Blue triangles, green dots, and red dots are the positions of tide gauge stations, GPSstations, and ocean bottom pressure gauges (PG1, PG2, and PG3), respectively, used in the inversions(Tables S1 and S2 in the auxiliary material). In the inset, yellow squares are the aftershocks with Mw ≥ 4in the 2 days after the main shock (U.S. Geological Survey); also indicated are the positions along the coastof Hokkaido where the runup measurements were higher than 4 m (Mabiro, Hamataiki, Horokayantou, andOikamanai [Tanioka et al., 2004a]).

1Auxiliary materials are available in the HTML. doi:10.1029/2009JB006665.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

2 of 12

possible, and in selecting only the recordings with a goodsignal‐to‐noise ratio. The selected records are sampled at1 min. Before using them in the inversion the tidal componentis removed by fitting it with a sum of three harmonics.Moreover, for each waveform, we choose a time windowincluding only the first oscillations of the tsunami wave tominimize the contribution of local effects on the tsunamisignal. Local effects, such as coastal reflections and resonanceof the bays could hide information on the seismic source andare particularly difficult to model.[8] The ground displacement associated to the earthquake

was recorded at the GPS stations of the GEONET network,operated by the Geographical Survey Institute of Japan(GSI). We use the GPS (Table S2 in auxiliary material andFigure 1) distributed over Hokkaido and northern Honshufor which the coseismic offsets were estimated by Larsonand Miyazaki [2008].[9] The vertical component of the coseismic displacement

has been recorded also on the seafloor by PGs, labeled asPG1 (depth 2218 m), PG2 (depth 2210 m) as well as with adepth sensor, labeled PG3 (depth 2540 m), that is includedin a conductivity‐temperature‐depth meter (CTD) sensor[Hirata et al., 2002]. In this work we use the vertical co-seismic static offsets at the PGs estimated by Mikada et al.[2006] (Table S2 in the auxiliary material).

3. Fault Parameterization

[10] The source area (Figure 1) is set on the basis of theMw ≥ 4 aftershock distribution in the 2 days after the mainshock. We set the strike of the fault (225°) according to thegeometry of the subducting plate proposed by Bird [2003];the dip angle of the plate boundary gradually changes withdepth (from 12° to ∼25° downdip, Table S3 in the auxiliarymaterial) [Katsumata et al., 2003; Hirata et al., 2003]. Therake angle (109°) is the best estimation of Yamanaka andKikuchi [2003]. We discretize the fault plane (210 × 150 km)dividing it into 35 subfaults with a size of 30 × 30 km each(Figure 1).[11] During a preliminary step of this work, we performed

some checkerboard tests with smaller subfault’s size (notshown), similar to those we describe in section 5. We con-cluded that, given the present data set and our inversionscheme, 30 × 30 km is very close to the minimum allowedsize for resolving details of the slip distribution.

4. Modeling and Green’s Functions

[12] We model the Green’s functions for each tide gaugeby using the nonlinear shallow water equations for thetsunami propagation [Satake, 2002]. Boundary conditionsare pure wave reflection at the solid boundary (coastlines)and full wave transmission at the open boundary (open sea).Equations are solved numerically by means of a finite dif-ference technique on a staggered grid [Mader, 2001]. Theinitial seawater elevation is assumed to be equal to thecoseismic vertical displacement of the sea bottom, whilethe initial velocity field is assumed to be identically zero. Thebathymetric grid for the computational domain (depicted inFigure 1) has 10 arc sec of spatial resolution and it was pro-vided by HOD.

[13] Green’s functions (horizontal and vertical coseismicdisplacements) at the GPS and PG stations, as well as theinitial condition for the tsunami propagation, are computedusing a layered Earth’s model by means of the PSGRN/PSCMP numerical code [Wang et al., 2006]. As inputparameters the PSGRN/PSCMP code needs P and S wavevelocities as well as density values for each of the layers.Here we use values from the vertical cross sections alongline F‐F′ as reported by Wang and Zhao [2005], adopting amodel with four layers, whose parameters are listed inTable S4 (see the auxiliary material).

5. Inversion Scheme and CheckerboardResolution Tests

[14] We solve the inverse problem using a particularimplementation of the simulated annealing technique calledthe “heat bath algorithm” [Rothman, 1986], following pre-vious studies using tsunami, GPS, and strong motion data[e.g., Piatanesi et al., 2007; Lorito et al., 2010]. For tsunamitime series, we use a cost function sensitive to both ampli-tude and phase matching [Spudich and Miller, 1990; Senand Stoffa, 1991]:

E mð Þ ¼XNk¼1

1�2Ptfti

uO tð ÞuS tð Þð ÞPtfti

u2O tð Þ þPtfti

u2S tð Þ

26664

37775k

ð1Þ

In equation (1) uO and uS are the observed and syntheticwaveforms, respectively, ti and tf are the lower and upperbounds of the time window, N is the number of records usedin the inversion, and m indicates the single model. For thegeodetic data set instead, we use a standard L2 norm as acost function to quantify the misfit between experimentaland synthetic data sets.[15] Regarding the joint inversion, the global cost function

is a weighted mean of the individual cost functions. Sinceeach cost function has a very different behavior dependingon its sensitivity to the variations of each parameter, it is notstraightforward to set the most appropriate weights. Forexample, during the synthetic tests described below, weverified that a progressive increase of the relative weightassigned to the tsunami data set with respect to the geodeticone results in a progressive loss of resolution for the sliponshore. On the other hand, by increasing the relativeweight of the geodetic data set results in a loss of resolutionon the slip offshore. Therefore, we empirically assignedsimilar weights to the two data sets, by means of severalsynthetic tests (analogous to those presented in sections 5.1,5.2, and 5.3).[16] We examine the resolving power on the slip distri-

bution of both data sets separately and of their combinationin a joint inversion, with several synthetic checkerboard tests[Piatanesi and Lorito, 2007; Lorito et al., 2008a, 2008b]. Thetarget model is a slip distribution featuring a checkerboardpattern, with slip values of 2 and 6 m alternating on adjacentsubfaults. Moreover, a priori information is introduced in themodel by imposing a range of variation to the slip value,which is allowed to vary from 0 to 10m on each subfault, witha 0.5 m step.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

3 of 12

Figure 2. Results of the checkerboard resolution tests. (a) Slip distribution of the target model. (b, c, d,and e) Slip distributions obtained inverting tsunami data, GPS data, geodetic data (GPS and PG data), andtsunami and geodetic data jointly, respectively. The star indicates the JMA’s epicenter of the 2003 Toka-chi‐oki earthquake. For each case, only the types of data used are plotted in the corresponding panel. Bluetriangles are tide gauges, green dots are GPS stations, and red dots are PGs. The tide gauges and GPSshown are only a fraction of those used for the inversions (see Figure 1). The blue ellipsis highlights afault portion where resolution is improved by adding PG data to GPS data (see text).

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

4 of 12

5.1. Checkerboard Test for Tsunami Data

[17] In order to mimic modeling uncertainties, the syn-thetic tsunami waveforms generated with the checkerboardmodel are corrupted by adding a Gaussian random noisewith a variance that is 10% of the clean waveform amplitudevariance [Sen and Stoffa, 1991; Ji et al., 2002].[18] The general pattern and the slip values of the target

model are fairly well recovered in the best model obtainedwith tsunami data (compare Figures 2a and 2b; see alsoTable S5 in the auxiliary material). However, a more carefulobservation reveals that slip values in the checkerboardpattern are almost perfectly recovered only on the portion ofthe fault far from the coast, whereas in the onshore zone theresiduals are larger (up to 1.5 m). This supports the generalconclusion that tsunami data constrain the offshore betterthan the onshore rupture features [Satake, 1993].

5.2. Checkerboard Test for Geodetic Data (GPSand PG)

[19] For GPS and PGs data, we add a Gaussian randomvalue to synthetic data sets, with a variance equal to thesquare of the standard deviation for each component (3 mmfor east, 4 mm for north, and 9 mm for vertical component[Larson and Miyazaki, 2008]).[20] With GPS data the assumed checkerboard pattern is

well reproduced only on the subfaults sufficiently close tothe stations and degrades farther from the coast (Figure 2cand Table S5 in the auxiliary material). GPS data alone

might be sufficient for constraining the offshore slip distri-bution only at a coarser scale (see the auxiliary materialFigure S1a). At the present scale (30 × 30 km) the slipdistribution could be recovered only with a hypotheticalconfiguration where the stations are immediately above andaround the whole fault zone (see the auxiliary materialFigure S1b).[21] Adding PG data helps to better constrain only the

fraction of the shallowest portion of the fault plane wherethe PGs are located (Figure 2d). However, geodetic ob-servations off Tokachi (Hokkaido) are still insufficient innumber and spatial density to well constrain the slip distri-bution offshore.

5.3. Checkerboard Test for Joint Inversion of Tsunamiand Geodetic Data

[22] A joint inversion combining the two kinds of datagives satisfactory results (Figure 2e and Table S5 in theauxiliary material), as the target model is well recoveredon the entire fault plane. The combination of tsunami andgeodetic data sets is then, in principle, efficient to constrainthe slip distribution for this earthquake, and it might be ingeneral a good strategy for retrieving the slip distribution oflarge tsunamigenic subduction zone earthquakes.[23] The uncertainties on the model parameters are esti-

mated, followingPiatanesi and Lorito [2007], bymeans of ana posteriori analysis of a subset of the ensemble of modelsexplored during the global search, for estimating the averagemodel and the uncertainties (see the auxiliary material).

Figure 3. Marginal distributions of the slip on each subfault for the synthetic joint inversion of Figure 2e(the order of the subfaults is given in the inset of Figure 1 and in Table S3 in the auxiliary material).Vertical blue dashed and red solid lines are the best and the average model values, respectively; horizontalblack solid bars are the standard deviations. Numerical values are reported in Table S5 (in the auxiliarymaterial).

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

5 of 12

Figure 3 shows that the best slip values are very close to theaverage ones and are well within the error bars. Then, the bestmodel well represents the subset of models that provide acomparable good fit to the data.[24] However, we note that the slip parameters 11–13, 16–

18, 21–23, and 26–28 are less resolved, as the correspond-ing distributions and standard deviations are relatively largerthan the others. Even if we still have sufficient control onthe results, as shown by the best model values, we havelarger uncertainties on this portion of the fault plane.

6. Results of Inversions: The 2003 Tokachi‐okiEarthquake

[25] We use both single and joint data sets inversions toestimate the slip distribution of the 2003 Tokachi‐okiearthquake. We assess by means of several tests (not shown)that the maximum slip value is well below 10 m, and thenwe use this value as a priori upper limit as in synthetic testsof section 5.

6.1. Inversion of Tsunami Data

[26] Tsunami data inversion indicates that there is a patchof high slip (6 m) located at about 30 km northwest of thehypocenter (Figure 4a, subfault 28, and Table S6 in theauxiliary material). The largest slip concentration is con-versely located in the deep part of the fault (between 60 and100 km NNW of the hypocenter), with slip values from 4 to6.5 m (subfaults 16, 17, 18, 21, and 22). Minor patches oflower slip (≤1.5 m) are mainly distributed in the SW portionof the fault. The differences between observed and predictedwaveforms are quantified by the global cost function valuecomputed by means of equation (1), and also by cost functionvalues for each tide gauge (Figure 5a).

6.2. Inversion of Geodetic Data

[27] The source model obtained with geodetic data(Figure 4b and Table S6 in the auxiliary material) featuresa stronger slip concentration (3.5–5 m) downdip from the

hypocenter (subfaults 21, 22, 23, 27, and 28). Differentlyfrom the tsunami data inversion, there is a maximum slipvalue of 7 m close to the hypocenter (subfault 24), and avery shallow patch with 4.5 m of slip at the southernmostfault corner (subfault 35). Slip is almost absent in the north-western part of the fault, and this feature is even morepronounced in this case than for the inversion of tsunamidata. Cost function values (the root mean square for thethree components of the displacement) give a measure of theresiduals between the observed and predicted coseismicdisplacements at the GPS and PG stations (Figure 5b).

6.3. Joint Inversion

[28] As the final step of the procedure, the joint inversionof tsunami and geodetic data is carried out. The sourcemodel (Figure 4c and Table S6 in the auxiliary material)shows an asperity (3.5–5.5m) located approximately betweenabout 30 and 80 km northwest of the hypocenter (subfaults22, 23, 27, and 28). A small and deeper patch of slip is presentat about 100 km northwest of the hypocenter (3 m, subfault21), and lower slip values (≤2 m) are distributed around themain asperity. After comparison of the synthetic tests(Figures 2b and 2d) to inversions of real data (Figures 4a,4b, and 4c) it is evident that the joint inversion modelcomprises features of the fault portions best resolved by thesingle inversions. We indeed observe that, in the jointinversion model, the pattern of slip distribution close to theGPS stations is similar to that obtained by inverting only thegeodetic data (Figures 4b and 4c, subfaults 21 and 26); onthe other hand, in the offshore part of the fault, althoughthree PGs are present, the slip distribution pattern is in fairagreement with the results obtained inverting only the tsu-nami data (Figures 4a and 4c, subfaults 24, 30, and 34).[29] Our best model features a seismic moment of 1.6 ×

1021 N m (using the depth‐dependent shear modulus valueslisted in Table S4 of the auxiliary material), correspondingto a magnitude Mw = 8.1.[30] The predicted tsunami waveforms agree very well

with the ones observed at the tide gauges (Figure 6a). We

Figure 4. Best model of the 2003 Tokachi‐oki earthquake slip distribution, obtained by inverting(a) tsunami data, (b) geodetic data, and (c) jointly tsunami and geodetic data. Symbols are as in Figure 2.Numerical slip values are reported in Table S6 (in the auxiliary material).

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

6 of 12

just notice almost rigid phase shifting in some cases (e.g., atKushiro, Tokachikou, and Muroran), probably due to sub-fault size or relatively poor bathymetry model. Additionally,also the predicted displacements fit very well the observedones (Figure 6b). Some discrepancies can probably beascribed to the size of the subfaults, or to the a priori fixedrake angle.[31] Best model cost functions for single data sets inver-

sions are slightly smaller than those for the joint inversion(Figures 5a and 5b), as in the latter more data have to be fitwith the same number of free parameters. However, thefact that the overall difference is small indicates an inter-consistency between the two data sets. Their consistency

also implies that it might be difficult to prefer one modelover another, as they roughly fit data equally well. Thewhole set of synthetic tests described above comes to therescue in this respect. Indeed, it is evident that the bestresolution over the fault plane is obtained only by the jointinversion (Figure 2e), whereas single inversions are not ableto resolve slip details everywhere (Figures 2b, 2c, and 2d).In other words, the joint inversion has conveniently reducedthe size of the null model space.[32] Figure 7 shows the marginal distributions for each

parameter. The best values are always within the error bars(i.e., the standard deviation). Best values can therefore beconsidered representatives of the subset of “good” models.

Figure 5. Goodness of fit for tsunami and geodetic data in terms of cost functions. (a) Tsunami data costfunctions. Black line is the global cost function value for tsunami data inversion of Figure 4a, and blackdots are the cost function values for single tide gauges (progressive numbers as in Table S1 the auxiliarymaterial). Red line and dots are the same quantities for the tsunami cost function in the joint inversion ofFigure 4c. (b) Same as Figure 5a except for the components of geodetic data inversion only (black) andfor the joint inversion (red). Corresponding inversion results are given in Figure 4b and 4c, respectively;progressive station numbers are as in Table S2 in the auxiliary material. The best model cost function forthe joint inversion is slightly larger than those for single data sets inversions because in the former, moredata have to be fit with the same number of free parameters.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

7 of 12

Figure 6. Comparison between the observed (black) and predicted (red) data sets. (a) Observed and pre-dicted tsunami waveforms. (b) Observed and predicted (left) GPS horizontal displacements and (right)GPS and PGs vertical displacements.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

8 of 12

As already observed in the distributions for the synthetic test(Figure 3), there is a subset of slip values that features rel-atively larger errors (e.g., subfaults 22, 23, 27, and 28).

7. Discussion

7.1. Comparison of the Slip Distribution Estimatedin This Study With Other Models

[33] The main asperity featured by our model is consistentat the first order with that imaged by some other modelsbased on teleseismic data [Yamanaka and Kikuchi, 2003;Horikawa, 2004], strong motion and GPS data [Koketsuet al., 2004], GPS data only [Miyazaki and Larson, 2008],and tsunami data [Tanioka et al., 2004b]. All of these modelsindeed feature a dominant slip patch with peak slip valuesranging from 4.5 to 7 m, while our model reaches a maximumslip of 5.5 m on two subfaults.[34] Conversely, there are important differences with the

models obtained by Robinson and Cheung [2010] (tele-seismic data), Aoi et al. [2008], Nozu and Irikura [2008](strong motion data), Yagi [2004] (teleseismic and strongmotion data jointly). Actually, Yagi [2004] finds three dif-ferent asperities, one surrounding the hypocenter and shiftedupdip with respect to our main patch, with a peak slip valueexceeding 5.5 m (Figure 8a). An almost colocated patch,with peak slip exceeding 11 m, appears in the model by Nozuand Irikura [2008]. The difference with Aoi et al. [2008] andRobinson and Cheung [2010] models is even more striking,as they both find significant slip updip of the hypocenter,

which is almost absent in any other model. The peak slip inthe shallow patch of Aoi et al. [2008] (Figure 8b) is morethan 6 m, whereas the corresponding patch of Robinson andCheung [2010] exceeds 12 m. However, Robinson andCheung [2010] claim that this shallow patch is not a robustfeature of their model. Furthermore, the azimuthal coverageoffered by strong motion stations is not optimal becausethey cover only the northwestern side of the fault area[Honda et al., 2004]. We agree with these conclusions, sincewe are convinced that if any such shallow and strong patch(6–12 m) existed for this earthquake, the ensuing tsunamiwould have been, at least locally, much more destructive.Source depth is indeed a first‐order parameter for tsunamigeneration efficiency, particularly for shallow depths. Accord-ingly, tsunami data inversions (in the present work, as wellas in the work by Tanioka et al. [2004b]) do not indicate anyshallow slip.[35] By comparing our slip distribution with the afterslip

distribution obtained byBaba et al. [2006], we observe that thepatch with the maximum amount of coseismic slip is enclosedby regions with large amount of afterslip (Figure 8c), con-sistently with the fact that the afterslip is expected to besmaller where the strain release is larger.

7.2. Some Implications to Next Large InterplateEarthquake off Tokachi

[36] Satake et al. [2005] divided the subduction zone in thesouthernmost Kuril Trench into three segments (Tokachi,Akkeshi, and Nemuro segments), and Hirata et al. [2009]

Figure 7. Marginal distributions of the slip on each subfault for the real case joint inversion of Figure 4c(the order of the subfaults is in the inset of Figure 1 and in Table S3 in the auxiliary material). Verticalblue dashed and red solid lines are the best and the average model values, respectively; horizontal blacksolid bars are the standard deviations. Numerical values are reported in Table S6 in the auxiliary material.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

9 of 12

suggested the presence of a distinct seismic gap just ontothe Akkeshi segment between the 2003 event and the 1973Nemuro‐oki earthquake (M7.7) [Tanioka et al., 2007](Figure 8c). The Akkeshi segment remained unbroken in2003, as already indicated in previous studies of the 2003tsunami source [e.g., Hirata et al., 2004; Tanioka et al.,2004b].[37] The coseismically ruptured area estimated in this

study is mainly concentrated downdip from the hypocenter

and, in comparison with Aoi et al. [2008] and Tanioka et al.[2004b], it includes only a small deep patch off Kushirowith a very low slip value (≤1 m). This patch of slip on theeastern end of the rupture area can then be considerednegligible for a large earthquake with a magnitude 8+ as the2003 Tokachi‐oki event. Moreover, the afterslip of the 2003event terminated just at the edge of the gap area (Figure 8c).Therefore our results confirm that the rupture does notextend on the plate interface off Akkeshi (Figure 8c).[38] The entire region (including the 2003 source zone

and the Akkeshi segment) is identified by Suwa et al. [2006]as one of the most strongly coupled regions in northeasternJapan, with a back slip accumulation rate of about 80 mm/yr. The main asperity in our source model corresponds to thearea of largest back slip [Hashimoto et al., 2009]. If we takeinto account the average slip (∼4 m) onto this area, we canroughly account for the amount of back slip accumulatedlocally since the 1952 event. A similar slip amount mightremain to be released on the Akkeshi segment; conversely,the accumulated stress might be accommodated by moderateearthquakes or in aseismic way [McCann et al., 1979]. TwosubsequentM∼7 earthquakes occurred in this zone (Figure 8c),in late 2004, likely releasing just a fraction of the accumu-lated stress. Therefore, the plate interface off Akkeshi shouldnot be an aseismic zone and it might be able to ruptureduring the next large interplate event along the Kuril Trench.

7.3. Possible Implications for a Tsunami ForecastSystem

[39] As highlighted in section 5 and as also pointed out byNishimura et al. [2005], offshore bottom pressure gaugescan help to better constrain the source characteristics ofan offshore earthquake. In case a tsunami is generated onlyfew minutes are available for launching an alert. A real‐timetsunami forecast system then must be based on the earliestavailable seismic data in order to estimate a first‐order tsu-nami threat level. Thus, a denser network of offshore sensorswould be a crucial improvement for a better rapid sourceestimation. Such a dense network might for example con-tribute to the current JMA’s tsunami early warning system,which is presently based on the maximum risk method fornear coastal events [Kamigaichi, 2008].

8. Conclusions

[40] We invert both tsunami (tide gauge records) andgeodetic data (GPS and PGs) to infer the slip distribution ofthe 25 September 2003 Tokachi‐oki earthquake. We per-

Figure 8. Comparison of the slip distribution of this workwith other models and with other earthquakes locations.(a) Black contours (1.2 m intervals) show the coseismic slipdistribution of the 2003 Tokachi‐oki by Yagi [2004];(b) black contours (2.0 m intervals) show the coseismic slipdistribution of the 2003 Tokachi‐oki by Aoi et al. [2008];and (c) black contours show the afterslip distribution of the2003 Tokachi‐oki earthquake [Baba et al., 2006], with acontour interval of 0.15 m; the slip distribution of the 1973Nemuro‐oki earthquake [Tanioka et al., 2007] is also shown;red dots represent two M∼7 earthquakes that occurred in thesuggested seismic gap off Akkeshi (blue ellipse).

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

10 of 12

form several synthetic tests in order to assess the sensitivityof the data to slip distribution details. We found that tsunamidata better retrieve the rupture features on the offshore faultportion, whereas geodetic data on the onshore zone. There-fore using jointly tsunami and geodetic data helps to betterimage the slip distribution of this large subduction zoneearthquake with respect to single data inversions. The sparsityof PG data contributes only slightly in resolving the slip onthe offshore portion of the fault plane. For this reason theinstallation of denser PG networks close to subduction zonesand their integration in real‐time tsunami forecast systems issuggested.[41] The slip distribution obtained by the joint inversion

features a maximum slip concentration (∼6 m) about 30–80 km northwest of the hypocenter, with a patch of slip (3 m)in the deep part of the source. The position of the mainasperities and the seismic moment (corresponding to a mag-nitudeMw 8.1) are in fair agreement with previous inversionsinvolving other kinds of geophysical data sets, with exceptionof those featuring shallow slip patches that are ruled out bythis study. Moreover, the slip does not extend on the plateinterface off Akkeshi. This supports a previous study [Hirataet al., 2009] that suggested the presence of a seismic gap inthe Akkeshi segment since the 1952 earthquake. Therefore,there is the possibility of a future large interplate event alongthat segment of the Kuril trench.

[42] Acknowledgments. We would like to thank the Japan Meteoro-logical Agency, the Hokkaido Regional Development Bureau of the Min-istry of Land, Infrastructure and Transport, and the Hydrographic andOceanographic Department of the Japan Coast Guard, for providing thedigital tide gauge records and the bathymetric data set.We also thank ClaudiaPiromallo, Stephen Monna and Giovanni Muscari for valuable suggestions,Zhi Wang for providing velocity profiles data, and Toshitaka Baba for pro-viding the postseismic distribution model. Some of the figures in this workwere drawn using GMT software [Wessel and Smith, 1995].

ReferencesAoi, S., K. Obara, S. Hori, K. Kasahara, and Y. Okada (2000), New strong‐motion observation network: KiK‐net, Eos Trans. AGU, 81(48), FallMeet. Suppl., Abstract S71A‐05.

Aoi, S., R. Honda, N. Morikawa, H. Sekiguchi, H. Suzuki, Y. Hayakawa,T. Kunugi, and H. Fujiwara (2008), 3D finite‐difference simulation oflong‐period ground motions for the 2003 Tokachi‐oki, Japan, earth-quake, J. Geophys. Res., 113, B07302, doi:10.1029/2007JB005452.

Baba, T., K. Hirata, T. Hori, and H. Sakaguchi (2006), Offshore geodeticdata conducive to the estimation of the afterslip distribution followingthe 2003 Tokachi‐oki earthquake, Earth Planet. Sci. Lett., 241, 281–292,doi:10.1016/j.epsl.2005.10.019.

Bird, P. (2003), An updated digital model of plate boundaries, Geochem.Geophys. Geosyst., 4(3), 1027, doi:10.1029/2001GC000252.

Hashimoto, C., N. Akemi, T. Sagiya, and M. Matsu’ura (2009), Interplateseismogenic zones along the Kuril‐Japan trench inferred from GPS datainversion, Nat. Geosci., 2, 141–144, doi:10.1038/ngeo421.

Hirata, K., et al. (2002), Real‐time geophysical measurements on the deepseafloor using submarine cable in the southern Kurile subduction zone,IEEE J. Oceanic Eng., 27, 170–181, doi:10.1109/JOE.2002.1002471.

Hirata, K., E. L. Geist, K. Satake, Y. Tanioka, and S. Yamaki (2003), Slipdistribution of the 1952 Tokachi‐oki earthquake (M8.1) along the KurilTrench deduced from tsunami waveform inversion, J. Geophys. Res.,108(B4), 2196, doi:10.1029/2002JB001976.

Hirata, K., Y. Tanioka, K. Satake, S. Yamaki, and E. L. Geist (2004), Thetsunami source area of the 2003 Tokachi‐oki earthquake estimated fromtsunami travel times and its relationship to the 1952 Tokachi‐oki earth-quake, Earth Planets Space, 56, 367–372.

Hirata, K., K. Satake, Y. Tanioka, and Y. Hasegawa (2009), Variable tsu-nami sources and seismic gaps in the southernmost Kuril Trench: A

review, Pure Appl. Geophys., 166, 77–96, doi:10.1007/s00024-008-0432-7.

Honda, R., S. Aoi, N. Morikawa, H. Sekiguchi, T. Kunugi, and H. Fujiwara(2004), Ground motion and rupture process of the 2003 Tokachi‐okiearthquake obtained by strong‐motion data of K‐NET and KiK‐net,Earth Planets Space, 56, 317–322.

Horikawa, H. (2004), Fault geometry and slip distribution of the 2003Tokachi‐oki earthquake as deduced from teleseismic body waves, EarthPlanets Space, 56, 1011–1017.

Ji, C., D. J. Wald, and D. V. Helmberger (2002), Source description ofthe 1999 Hector Mine, California earthquake; Part I: Wavelet domaininversion theory and resolution analysis, Bull. Seismol. Soc. Am., 92(4),1192–1207, doi:10.1785/0120000916.

Kamigaichi, O. (2008), Tsunami forecasting and warning, in Encyclopediaof Complexity and System Science, vol. 10, pp. 9592–9618, Springer,New York.

Katsumata, K., N. Wada, and M. Kasahara (2003), Newly imaged shape ofthe deep seismic zone within the subducting Pacific plate beneath the Hok-kaido corner, Japan‐Kurile arc‐arc junction, J. Geophys. Res., 108(B12),2565, doi:10.1029/2002JB002175.

Kinoshita, S. (1998), Kyoshin net (k‐net), Seismol. Res. Lett., 69, 309–332.Koketsu, K., K. Hikima, S. Miyazaki, and S. Ide (2004), Joint inversion ofstrong motion and geodetic data for the source process of the 2003Tokachi‐oki, Hokkaido, earthquake, Earth Planets Space, 56, 329–334.

Larson, K. M., and S. Miyazaki (2008), Resolving static offsets from high‐rate GPS data: The 2003 Tokachi‐oki earthquake, Earth Planets Space,60, 801–808.

Lorito, S., F. Romano, A. Piatanesi, and E. Boschi (2008a), Source processof the September 12, 2007, Mw 8.4 southern Sumatra earthquake fromtsunami tide gauge record inversion, Geophys. Res. Lett., 35, L02310,doi:10.1029/2007GL032661.

Lorito, S., A. Piatanesi, and A. Lomax (2008b), Rupture process of the 18April 1906 California earthquake from near‐field tsunami waveforminversion, Bull. Seismol. Soc. Am., 98(2), 832–845, doi:10.1785/0120060412.

Lorito, S., A. Piatanesi, V. Cannelli, F. Romano, and D. Melini (2010),Kinematics and source zone properties of the 2004 Sumatra‐Andamanearthquake and tsunami: Nonlinear joint inversion of tide‐gage, satellitealtimetry and GPS data, J. Geophys. Res., 115, B02304, doi:10.1029/2008JB005974.

Mader, C. L. (2001), Numerical Modeling of Water Waves, CRC Press,Boca Raton, Fla.

McCann, W. R., S. P. Nishenko, L. R. Sykes, and J. Krause (1979), Seismicgap and plate tectonics: Seismic potential for major boundaries, Pure Appl.Geophys., 117, 1082–1147, doi:10.1007/BF00876211.

Mikada, H., et al. (2006), New discoveries in dynamics of anM8 earthquake‐Phenomena and their implications at the 2003 Tokachi earthquake using along term monitoring cabled observatory, Tectonophysics, 426, 95–105,doi:10.1016/j.tecto.2006.02.021.

Miyazaki, S., and K. M. Larson (2008), Coseismic and early postseismicslip for the 2003 Tokachi‐oki earthquake sequence inferred from GPSdata, Geophys. Res. Lett., 35, L04302, doi:10.1029/2007GL032309.

Nishimura, S., M. Ando, and K. Tadokoro (2005), An application ofnumerical simulation techniques to improve the resolution of offshorefault kinematics using seafloor geodetic methods, Phys. Earth Planet.Inter., 151, 181–193, doi:10.1016/j.pepi.2005.03.002.

Nozu, A., and K. Irikura (2008), Strong‐motion generation areas of a greatsubduction‐zone earthquake: Waveform inversion with empiricalGreen’s functions for the 2003 Tokachi‐oki earthquake, Bull. Seismol.Soc. Am., 98(1), 180–197, doi:10.1785/0120060183.

Piatanesi, A., and S. Lorito (2007), Rupture process of the 2004 Sumatra‐Andaman earthquake from tsunami waveform inversion, Bull. Seismol.Soc. Am., 97(1A), S223–S231, doi:10.1785/0120050627.

Piatanesi, A., P. Heinrich, and S. Tinti (1999), The October 4, 1994 Shikotan(Kurile Islands) tsunamigenic earthquake: An open problem on the sourcemechanism, Pure Appl. Geophys. , 154, 555–574, doi:10.1007/s000240050244.

Piatanesi, A., A. Cirella, P. Spudich, and M. Cocco (2007), A global searchinversion for earthquake kinematic rupture history: Application to the2000 Western Tottori, Japan earthquake, J. Geophys. Res., 112,B07314, doi:10.1029/2006JB004821.

Robinson, D. P., and L. T. Cheung (2010), Source process of the Mw 8.3,2003 Tokachi‐Oki, Japan earthquake and its aftershocks, Geophys. J. Int.,181(1), 334–342, doi:10.1111/j.1365-246X.2010.04513.x.

Rothman, D. (1986), Automatic estimation of large residual statics correc-tions, Geophysics, 51, 332–346, doi:10.1190/1.1442092.

Satake, K. (1993), Depth distribution of coseismic slip along the NankaiTrough, Japan, from joint inversion of geodetic and tsunami data, J. Geo-phys. Res., 98(B3), 4553–4565, doi:10.1029/92JB01553.

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

11 of 12

Satake, K. (2002), Tsunamis, in International Handbook of Earthquakeand Engineering Seismology, Int. Geophys. Ser., vol. 81(A), edited byW. H. K. Lee et al., pp. 437–451, Academic, London.

Satake, K., F. Nanayama, S. Yamaki, Y. Tanioka, and K. Hirata (2005),Variability among tsunami sources in the 17th‐21st centuries along thesouthern Kuril Trench, in Tsunami Research: Case Studies and RecentDevelopments, Adv. Nat. Technol. Hazards Res., vol. 23 edited byK. Satake, pp. 157–170, Springer, Dordrecht, Netherlands.

Sen, M., and P. L. Stoffa (1991), Nonlinear one‐dimensional seismic wave-form inversion using simulated annealing, Geophysics, 56, 1624–1638,doi:10.1190/1.1442973.

Spudich, P., and D. P. Miller (1990), Seismic site effects and the spatialinterpolation of earthquake seismograms: Results using aftershocks ofthe 1986 North Palm Springs, California earthquake, Bull. Seismol.Soc. Am., 80, 1504–1532.

Suwa, Y., S. Miura, A. Hasegawa, T. Sato, and K. Tachibana (2006), Inter-plate coupling beneath NE Japan inferred from three‐dimensional dis-placement field, J. Geophys. Res., 111, B04402, doi:10.1029/2004JB003203.

Tanioka, Y., et al. (2004a), Tsunami run‐up heights of the 2003 Tokachi‐okiearthquake, Earth Planets Space, 56, 359–365.

Tanioka, Y., K. Hirata, R. Hino, and T. Kanazawa (2004b), Slip distribu-tion of the 2003 Tokachi‐oki earthquake estimated from tsunami wave-form inversion, Earth Planets Space, 56, 373–376.

Tanioka, Y., K. Satake, and K. Hirata (2007), Recurrence of recent largeearthquakes along the southernmost Kurile‐Kamchatka subduction zone,in Volcanism and Subduction: The Kamchatka Region, Geophys.Monogr. Ser., vol. 172, edited by J. Eichelberger et al., pp. 145–152,AGU, Washington, D. C.

Utsu, T. (1999), Seismicity Studies: A Comprehensive Review (in Japanese),876 pp., Univ. of Tokyo Press, Tokyo.

Wang, R., F. L. Martin, and F. Roth (2006), PSGRN/PSCMP—A new codefor calculating co‐ and post‐seismic deformation, geoid and gravitychanges based on the viscoelastic‐gravitational dislocation theory, Comput.Geosci., 32, 527–541, doi:10.1016/j.cageo.2005.08.006.

Wang, Z., and D. Zhao (2005), Seismic imaging of the entire arc of Tohokuand Hokkaido in Japan using P‐wave, S‐wave and sP depth‐phase data,Phys. Earth Planet. Inter., 152, 144–162, doi:10.1016/j.pepi.2005.06.010.

Watanabe, T., H. Takahashi, M. Ichiyanagi, M. Okayama, M. Takada,R. Otsuka, K. Hirata, S. Morita, M. Kasahara, and H. Mikada (2006),Seismological monitoring on the 2003 Tokachi‐oki earthquake, derivedfrom off Kushiro permanent cabled OBSs and land‐based observations,Tectonophysics, 426, 107–118, doi:10.1016/j.tecto.2006.02.016.

Wessel, P., and W. H. F. Smith (1995), New version of the Generic Map-ping Tools released, Eos Trans. AGU, 76, 329, doi:10.1029/95EO00198.

Yagi, Y. (2004), Source rupture process of the 2003 Tokachi‐oki earth-quake determined by joint inversion of teleseismic body wave and strongground motion data, Earth Planets Space, 56, 311–316.

Yamanaka, Y., and M. Kikuchi (2003), Source process of the recurrentTokachi‐oki earthquake on September 26, 2003, inferred from teleseismicbody waves, Earth Planets Space, 55, e21–e24.

K. Hirata, Seismology and Volcanology Research Department,Meteorological Research Institute, Japan Meteorological Agency, Tsukuba305‐0052, Japan.S. Lorito, A. Piatanesi, and F. Romano, Sezione di Sismologia e

Tettonofisica, Istituto Nazionale di Geofisica e Vulcanologia, Via diVigna Murata 605, I‐00143 Rome, Italy. (fabrizio.romano@ingv.it)

ROMANO ET AL.: TOKACHI‐OKI 2003 JOINT SLIP INVERSION B11313B11313

12 of 12