We reconstructed new migrated images from autocorrelation functions of ambient noise at seismic stations in northeastern Japan. During the analysis period, the autocorrelation functions, having a frequency range of 0.5–2 Hz, showed some coherent signals at large lapse times of 10–50 s as well as at small lapse times of less than 10 s. These signals were also coherent in space along cross sections across the Japan subduction zone. The migrated images show seismic scatterers within the subducting Pacific Slab at the depth range of 50–150 km. The distribution of the scatterers within the subducting slab is comparable to the double seismic zone in the subduction zone; in particular, the thickness of the distribution of the seismic scatterers is consistent with that of two planes of the double seismic zone. This indicates inhomogeneities in the seismic velocity structure around the upper and lower planes of the seismicity. The new migrated images suggest that seismic scatterers, such as fluid from dehydration reactions of the hydrated metamorphosed mantle, exist in the lower as well as upper plane of the seismicity in the double seismic zone; this fluid may act as a scatterer within the subducting slab.
 Seismic reflection surveys are widely used in exploration geophysics to image shallow as well as deep subsurface structures. To carry out an effective survey, a controlled seismic source having the maximum possible energy is required. However, the various sources used for such surveys, such as explosives or seismic vibrators, all suffer from different limitations.
 As an alternative, therefore, the use of seismic interferometry has been proposed. In this technique, Green's functions of seismic waves between two points are obtained by using the correlation functions of ambient noise or coda waves [e.g., Campillo and Paul, 2003; Schuster et al., 2003; Wapenaar et al., 2004]. This technique can be used to investigate Earth's subsurface in detail through an analysis of the velocity boundary of seismic waves. It is well known that surface waves can be retrieved by cross-correlating the ambient seismic noise generated by multiple scattered waves [Shapiro and Campillo, 2004]. Some studies observed body waves reflected from subsurface interfaces from the cross-correlogram of P- and S-coda waves [Draganov et al., 2007; Abe et al., 2007]. Others retrieved reflections from the correlations of ambient noise [Poli et al., 2012; Ito et al., 2012]. For example, Poli et al. identified body waves with a frequency of 0.5–2 Hz emerging from the cross-correlation of ambient noise for inter-station distances of up to 550 km and showed that these waves can be interpreted asP and S waves reflected from the Moho. Furthermore, Ito et al. obtained a migrated image by calculating autocorrelations as zero-offset shot traces of ambient noise and revealed a reflected wedge mantle structure and relatively transparent subducting slab in the Japan subduction zone.
 The Japan subduction zone marks the zone along which the Pacific plate subducts beneath the landward plate (Figure 1), generating not only interplate earthquakes but also intraplate and intraslab events. These seismicities form a typical double seismic zone with two parallel planes of seismicity that are 30 km apart in the 50–200 km depth range [Hasegawa et al., 1978]. Upper plane earthquakes occur within the subducting oceanic crust [Kita et al., 2006], whereas lower plane earthquakes occur within the subducting oceanic mantle. High-resolution seismic tomography studies of the subducting Pacific Plate beneath northeastern Japan [Tsuji et al., 2008; Nakajima et al., 2009a, 2009b] that were carried out in recent years revealed a low seismic velocity with Vp/Vs below 1.73 in the shallow oceanic mantle. This was attributed to serpentine dehydration, which suggests that the lower plane seismicity is caused by dehydration embrittlement [Peacock, 2001]. The geometry of the upper boundary of the subducting plate beneath northeastern Japan has been mapped by converted and reflected waves [Matsuzawa et al., 1986; Zhao et al., 1997; Nakajima et al., 2002] and receiver function analysis [Chen et al., 2005]. The upper surface of the subducting plate has been well imaged by local tomography using the first P- and S- arrival times as well as the Moho discontinuity of the continental plate [Zhao et al., 1990; Nakajima et al., 2001]; the upper surface of the subducting plate is well comparable to the upper plane of the double seismic zone.
 Here, we apply the autocorrelation technique using the ambient noise observed using a highly dense seismic network to extract scattered waves from the subducting slab from a slant migrated image that is considered a tilting structure related to the subducting slab. We also discuss possible reflectors related to the double seismic zone within the subducting slab.
 We calculated the autocorrelation functions (ACFs) of ambient noise with a time window length of 120 s at seismic stations located at Tohoku University and the Hi-net stations operated by the National Research Institute for Earth Science and Disaster Prevention (NIED).Figure 1shows the location of the seismic stations used in this study. All of the ACFs were calculated from the continuous data of the Z-component at each station. Almost all the seismometers located at the Hi-net stations were installed at the bottom of boreholes with depths of 100 m [Okada et al., 2004]; some were installed at the bottom of deeper boreholes with depths of 2000 m, but these were not used in our study because of the effects of the reflection waves at these depths from the surface. The natural frequency of the seismometers at the Hi-net stations was 1 Hz. All the short-period seismometers with a natural frequency of 1 or 2 Hz at the Tohoku University seismic stations were installed on the surface. All the continuous data were recorded with a 100-Hz sampling frequency. In addition, the broadband seismometers of Tohoku University were installed in tunnels near the surface. The analysis in our study was carried out from June 2006 to the end of March 2007.
3. Data processing
3.1. Autocorrelation of Continuous Records
 We express the autocorrelation as a function of the lapse time τ for velocity waveforms of continuous seismic data v(t) in the time domain as
where Tw is the time window length for the calculation of the correlation coefficient at lapse time τ.Filtered 1-h continuous data in the frequency range of 0.5–2 Hz is used to calculate the autocorrelation using a 1-bit correlation technique [Campillo and Paul, 2003]. By taking the ensemble average of ACFs over 30 days, 30-day ACFs are calculated for more than 300 days at each station.
3.2. Depth-Migrated Image
 The ACFs from the vertical components should be reconstructed either from body waves emitted and striking near-vertically at the station or from surface waves passing through and returning to the station. Here we assume a P-wave velocity structure model and PP reflections between the stations and the reflectors, and we compute a depth-migrated image from the ACFs as zero-offset shot traces as done byIto et al. . By assuming that the ACF amplitude at lapse time τ at a station i is obtained from those of the reflected waves, we calculate the normalized stacked amplitude (NSA) at each grid node at location x in a 3D volume as given below.
where τidenotes the two-way travel time between grid nodex and station i at xi as defined by τi = 2·Tp(x; xi); Tp, the P-wave travel time from the grid node to the station; andN, the number of ACFs used to calculate the NSA. The x-axis of a Cartesian coordinate system in a positive orientation is chosen along each profile of A-A′, B-B′, C-C′ and D-D′. We calculated the NSA at each grid node by assuming the simple 1D velocity model that is adopted in the NIED F-net moment tensor analysis (Figure S1 in theauxiliary material) [Fukuyama et al., 1998] and PP reflections between the stations and the reflectors.
 All NSAs were calculated from the combination of the data obtained at the stations and the grid nodes. All the reflected waves from the grid nodes to the stations have an incident angle of 5°–25°, and the back azimuth of a grid node as a scatterer is within 20° from the azimuth of the maximum inclined angle of the subducting Pacific Plate in order to image a slanted structure related to the subducting slab while considering the incident angle to the station and take-off angle from the grid node. All of the grid nodes are arranged within intervals of 2 km in space.
 Finally, we calculate 100-km-wide cross-sections of the migrated images along a survey line; here, a stacked amplitude at a point (x,z) is calculated from the NSA stacked along the x-direction as defined by the following expression:
where x, y, and z correspond to the length, width, and depth directions, respectively, of the cross section. M denotes the number of grids along y. As follows, we chose the stacked width along the x-direction from −50 to 50 km.
 The ACFs obtained for the analysis period show some coherent signals in time with relatively large amplitude as in the case of Ito et al. . Figure 2shows the 30-day ACFs along the four cross sections of the chosen survey lines during the analysis period. The ACFs show signals at both large lapse times of 10 to 50 s as well as small lapse times of 10 s. The latter correspond to the depth of the continental Moho discontinuity. The former likely correspond to the depth of the mantle wedge or the upper surface of the subducting Pacific Plate; these signals are identified clearly at stations located in the back-arc side. In fact, these signals frequently appear at the back-arc side, suggesting that some reflectors probably exist between the Moho and the plate boundary. In contrast, weak signals with lapse times of 20–30 s corresponding to the depth of the plate boundary are detected at some stations located in the fore-arc side. These results suggest that the ACFs correspond to Green's functions, thereby indicating a tilting seismic structure related to the subducting Pacific slab.
Figure 3 shows the new migrated images reconstructed from the ACFs. We compared the migrated images with the Moho discontinuity and plate boundary between the subducting Pacific Plate and the overriding landward plate, which have been extensively investigated by Zhao et al. [1990, 1992] and Nakajima et al. [2001, 2002]. The large NSAs or scatterers near the Moho are not so clear. Taking into account the stacked number of migrated images shown in Figure S2, this can be attributed to the relative lack of sufficient stacked numbers required to construct the migrated images. On the other hand, the scatterers between the Moho and the upper surface of the subducting slab were well imaged; the migrated image showed a reflected wedge mantle structure and relatively transparent subducting slab. This wedge mantle structure was the same as that reported by Ito et al. .
 At a depth range of 50–100 km, the slanted distribution of the scatterers near the plate boundary between the subducting Pacific Plate and wedge mantle is well imaged. This probably shows reflectors or scatterers along or immediately below the plate boundary. The slanted structure along the plate boundary is not clearly identified in the other migrated images with other conditions for incident angles of 25°–45° and 35°–55°, although a similar structure is somewhat imaged for an incident angle of 15°–35° (Figure S3). This result suggests that reflectors that are slanted in the same manner in the slab and in the mantle wedge are not indicative of a bias in the migration.
 The scatterers are also distributed within the subducting slab, although this slab is more transparent than the overriding mantle wedge. Scatterers with a thickness of 30–50 km are distributed from the uppermost part of the subducting slab. Scatterers are observed to a depth range of 50 and 150 km within the subducting slab; in particular, double scatterers with two parallel planes of reflectors that are 30–50 km apart are observed at the depth range of 50–150 km along the C-C′ profile inFigure 3. The thickness of the scatterer's structures is consistent with that of the distribution of the double seismic zone with parallel planes of seismicity. The bottom of the scatterers within the slab dramatically turns into a transparent structure along the C-C′ profile as well as the A-A′ and B-B′ profiles, although the D-D′ profile shows a more transparent slab structure than do the other profiles.
5. Discussion and Conclusions
 The new migrated images shown in Figure 3show the scatterers slanted along or immediately below the uppermost part of the subducting slab. The scatterers within the subducting slab suddenly disappear at a depth of 100 km, as does the seismicity along the upper plane of the double seismic zone, although they are well identified to a depth of 120 km in the A-A′ profile and to a depth of 100 km in the B-B′ and C-C′ profiles.Tsuji et al. used a seismic velocity tomography analysis to demonstrate that a remarkable low-velocity zone with a thickness of 10 km is imaged at the uppermost part of the slab in the Japan subduction zone; this zone gradually disappears at depths of 70–90 km, which is nearly identical to the B-B′ profile in this study, suggesting the occurrence of intensive dehydration reactions there. Our migrated images provide further evidence, in which the scatterers slanted along the plate boundary in the slab and in the mantle wedge probably correspond to the fluid distribution related to the dehydration reaction.
 The scatterers above 10–20 km away from the upper surface of the subducting slab are well identified from a depth of more than 80–100 km within the back-arc-side wedge mantle, especially along the A-A′ and B-B′ profiles.Kawakatsu and Watada conducted a receiver function analysis to demonstrate that a low-velocity layer exists above the slab from a depth of 80 to 130 km and that it extends to depths of at least 100–120 km along the slab along a cross section that is nearly identical to the B-B′ profile in this study; the low-velocity zone above the slab corresponds to a hydrous layer. Our images probably correspond to the upper surface of a fully hydrous layer, as was also reported byIto et al. .
 Our migrated image also showed possible scatterers corresponding to the depth of the lower as well as the upper plane of the double seismic zone. Double seismic zones with two parallel planes of seismicity have been observed at the depth range of 50–200 km in several subduction zones. However, this is not sufficient to account for the occurrence of earthquakes at depths greater than 30 km, because the lithostatic pressure becomes too high for any brittle fracture or frictional slip to be possible [Frohlich, 1994; Kirby, 1995]. The hypothesis of dehydration embrittlement of metamorphosed oceanic crust and mantle in the subducting slab is one of the possible key processes controlling the occurrence of lower as well as upper plane seismicity [Green and Burnley, 1989; Kirby et al., 1991; Yamasaki and Seno, 2003]. Peacock  showed that lower plane earthquakes occur at 550–800°C at a depth of 100 km and at 350–600°C at a depth of 160 km; these conditions coincide with the dehydration reaction from antigorite (serpentine) to forsterite, enstatite, and H2O. If the hypothesis is true, a considerable amount of fluid should exist around the lower plane of the double seismic zone because of the dehydration of the metamorphosed mantle within the slab. These fluids may act as scatterers near and around the lower plane seismicity.
 We reconstructed the migrated images from the autocorrelation for ambient noise in northeastern Japan. The migrated images at the depth range of 50–150 km show the seismic scatterers within the subducting slab. The thickness of the distribution of the scatterers was consistent with that of the double seismic zone, suggesting inhomogeneities in the seismic velocity structure around the upper and lower planes of the double seismic zone because of the fluid formed by the dehydration reaction.
 We thank Satoshi Kaneshima and an anonymous reviewer for their comments. We also thank Akira Hasegawa and Stephen Kirby for the helpful discussions. This study was also supported by JSPS KAKENHI (22740288). We used the hypocenter catalogue compiled by the Japan Meteorological Agency. The figures were prepared using the Generic Mapping Tool [Wessel and Smith, 1998].