Our initial data consist of P and S wave arrival times of local earthquakes recorded by the SCSN, compiled by the Southern California Earthquake Data Center and obtained through the Seismogram Transfer Program (http://www.data.scec.org/research-tools/stp-index.html). In the study area (blue box in Fig. ), SCSN data analysts have picked the phase data (first P and S arrivals) of nearly 30 000 earthquakes with magnitudes between 2.0 and 4.0 occurring during a period from January 1992 to November 2013. Since it is very computationally intensive and also unnecessary to include all these events, we only choose a small subset of them for our tomographic inversion.

To ensure that the chosen seismic data illuminate the study region well, events and corresponding phase records are carefully selected based on the following six criteria: (1) to guarantee the quality of seismic data and validity of point source assumption for forward modeling, the magnitudes of the selected events should be within the range [2.0,4.0]; (2) to reduce the influence of mislocation errors on tomographic inversion, we only choose events with more than 20 P and more than 10 S arrivals; (3) the focal depth of each chosen event is greater than 3.0 km; (4) to ensure that picking errors of selected phase data are within an acceptable range, the misfit between the observed arrival time and the synthetic arrival time in the 1-D reference model (discussed later) is required to be less than 1.0 s for P wave or 1.5 s for S wave; (5) to save on computation, we only use seismic records whose epicentral distances are less than 100 km; (6) to avoid event clustering and keep a uniform distribution of hypocenter locations, we divide the Landers earthquake source area (the blue box in Fig. ) into 2 km × 2 km × 2 km blocks and only choose one event in each block that is recorded by the maximal number of stations if it exists. As a result, our selected data set includes 78 523 first P wave and 46 999 first S wave arrival times recorded by 275 SCSN stations (Fig. b) for 2041 local earthquakes (Fig. a).

The discrete form of tomographic Eq. () requires model parameterization. We first need to define the forward modeling grid for the calculation of travel-time kernel K(x;xr,xs) in Eq. (2). Since in this study the travel-time kernel K(x;xr,xs) is computed on the vertical plane passing through the source xs and the receiver xr by numerically solving the two acoustic wave Eqs. () and () using a finite-difference scheme (i.e., the high-order central difference method presented in the appendix of paper I), the forward modeling grid should be designed to suit 2-D finite-difference calculations. Usually, for a finite-difference calculation, the computational domain is divided into a uniform grid where the grid size is determined by the velocity, dominant frequency of seismic wave, and stability condition of the numerical scheme.

We first define m0 as a 1-D layered velocity model that contains five layers separated by two velocity boundaries at 2.0 km and 5.5 km, the Conrad discontinuity (16 km), and an averaged flat Moho (29.2 km) . In each layer, the velocity structure is homogeneous and the corresponding P and S wave velocities are shown in Table . For this 1-D layered model, the arrival times of the direct P and S waves, head waves refracted from the velocity boundary at the depth of 5.5 km observed at epicentral distances > 40–50 km, head wave (P*, S*) refracted from the Conrad discontinuity when the epicentral distances are in the range of 90–140 km, and head waves (Pn, Sn) from the Moho when the epicentral distances are greater than 140–150 km can be easily calculated according to the geometrical ray theory . Accordingly, the synthetic first P and S arrival times can be determined for each source–receiver pair based on its epicentral distance for the velocity model m0. However, the undulated Moho of southern California region has large lateral depth variation and strong influence on seismic wave propagation , and has considerable effects on the tomographic images of the lower crust and the uppermost mantle e.g.,. Therefore, for this study, we take into account the variation of Moho topography, and introduce a velocity model m1 that differs m0 by adding an undulated Moho obtained from receiver functions by as the starting model. The synthetic travel times of the first P and S waves in m1 can be calculated by the combined ray and cross-correlation technique discussed in Paper I. Once the synthetic arriving times for m1 and the observed arrival times picked from data are available, velocity structures can be updated from m1 based on the WETST technique.

For the 2-D finite-difference forward modeling, we choose a Gaussian wavelet as the source time function f(t) in Eq. () f(t)=A2π2f02t-1.2f02-1exp⁡-π2f02t-1.2f02, where A is the amplitude and f0 is the dominant frequency. The frequency spectrum of the source time function Eq. () is mainly concentrated within [0,2.5f0]. For example, the spectrum (shown in Fig. b) for a f(t) with unit amplitude A=1.0 and dominant frequency f0=2.0 Hz (Fig. a) has significant values between 0.0 Hz and 5.0 Hz. Correspondingly for consistency, data traces need to be filtered between the frequency range of [0,2.5f0] for the picking of observed travel times. We specifically denote the observed travel times picked on band-pass filtered seismograms as Tobs,f. Since seismic waves filtered at different frequencies have different sensitivity to heterogeneous structures, arrival-time Tobs,f is not necessarily equal to Tobs obtained from the travel-time data catalog (such as the SCSN). We relate the two arrival times using the formula Tobs,f=Tobs+δtf. Fortunately, δtfs are found to be very small and negligible in this study. In detail, we first choose the dominant frequencies f0=2.0 Hz for P waves and f0=1.2 Hz for S waves, since the dominant parts of the seismic energy are around these frequencies for moderate crustal earthquakes e.g.,. The wavelengths of the P wave are approximately equal to those of the S wave in the same layers. To explore the properties of the arrival-time difference δtf in Eq. (), a Butterworth filter between 0.001 Hz and 5.0 Hz is applied to more than 50 arbitrarily selected P wave seismograms recorded for 10 earthquakes with magnitudes between 2.08 and 3.99. Figure  shows three such examples of Tobs and Tobs,f picked on raw and filtered seismograms. In all our selected examples we find that the differences δtf are generally smaller than 0.08 s, which account for the combined effect of finite-frequency measurements, noise, and picking inaccuracy. For a regional P wave tomography as in this study, δtf less than 0.08 s has very limited effect on the final images and can be safely viewed as noise, which will also be confirmed in the checkerboard resolution tests shown in the Supplement (Figs. S1 and S2). Similarly, δtf can also be ignored for the S wave seismograms. Therefore, we would rather use the existing SCSN catalog of Tobs than hand-picking large number of Tobs,f for this tomographic study. In addition, the spacing of the uniform forward modeling grid is Δx=Δz=0.2 km. The time steps are chosen to be Δt=0.0025 s for P wave simulations and Δt=0.004 s for S wave modelings. These parameters guarantee the stability condition of the high-order central difference method .

One of the main purposes of forward modeling is to compute the sensitivity kernel as in Eq. (2). Prior to that, we need to determine the time window for the first arriving P-phase or S-phase. Since the onset time Tsyn of a specific phase in an iterative model can be calculated by using the combined ray and cross-correlation method and the signal length (dependant on the source time Eq. 5) is about twice the dominant period, the time window for the synthetic seismic phase is chosen to be [Tsyn,Tsyn+2.0/f0]. For the 1-D layered model with an undulated Moho (m1) and at epicentral distances less than 100 km, the first P or S wave arrival of a crustal earthquake (depth greater than 3.0 km) should be either the direct phase Pg (Sg), head waves refracted from the velocity boundary at the depth of 5.5 km, or head waves P* (S*) refracted from the Conrad discontinuity depending on the epicentral distance. Accordingly, the travel-time sensitivity kernels for the first arrivals also have different spatial variations. Figures a, b and a, b show two typical sensitivity kernels for the velocity model m1. For example, the kernels of Pg and Sg waves at a distance of 3.75 km for an earthquake at the depth 3.14 km (Fig. a and b) clearly display 2-D cigar shapes e.g.,. On the synthetic seismograms (Fig. c and d), the direct Pg (Sg) and the reflected phase Pr (Sr) from the velocity boundary at 5.5 km are distinguishable and almost totally separated, which enables us to separate the direct arrivals and calculate their kernels (Fig. a and b). The negative kernel values in the first Fresnel zone indicate that a velocity decrease is required to delay the synthetic arrival time Tsyn. However, for the records of 2041 selected crustal events, it is only possible to separate the first arrival from its coda waves on a very small fraction of synthetic seismograms. For many synthetic seismograms, the first arrivals are closely followed or even overlapped by other phases. For example, on the synthetic seismograms generated by the same crustal earthquake but recorded at a distance of 87.24 km (Fig. c and d), the first arrival Ph (Sh) refracted from the velocity boundary at 5.5 km depth is sequentially overlapped by the direct arrival Pg (Sg), the Conrad refracted phase P* (S*), and the reflected wave Pr (Sr) from the velocity boundary at 5.5 km depth. It is difficult to separate these phases because the later phases are in the time window of the first P wave (S wave) arrival. Therefore, the computed sensitivity kernels (Fig. a and b) have significant values around the traveling paths of all the phases that arrive within the first-arrival windows. This feature is helpful for resolving multipathing problems which are common for complex velocity structures . Every P wave/S wave travel-time sensitivity kernel is also smoothed out by a Gaussian function with the scaling length chosen to be the minimum P wave/S wave wavelength in the starting model m1 . Additionally, the sensitivity kernel K(x;xr,xs) and the relative velocity perturbation δc(x)/c(x) are bilinearly interpolated on the forward modeling grid in this study.

Once all the travel-time sensitivity kernels are calculated, smoothed, and interpolated, we can invert for the relative velocity perturbation field δc(x)/c(x). As discussed in paper I, the relative velocity perturbation field δc(x)/c(x) at each forward modeling grid node is linearly interpolated by its values at the eight neighboring inversion grid nodes. Based on the data distribution as shown in Fig. a, we setup the inversion grid in the study area (Fig. b) with a horizontal grid spacing of 0.12∘ at the central potion and 0.15∘ near the edges (Fig. b), and seven vertical layers located at the depths of 1, 5, 10, 15, 21, 28, and 40 km. The spacing of the chosen inversion grid is much larger than that of the forward modeling grid. Additionally, the minimum wavelengths of both P and S waves are approximately half of the minimum inversion grid size. Generally speaking, at least four grid nodes per wavelength are needed to fully capture the seismic wavefield with a finite-difference forward modeling method e.g.,. If an inverse algorithm has a resolving ability at the scale of the wavelength λ, such as that of full waveform inversion methods , the grid spacing of the inversion grid should be at least 4 times that of the forward modeling grid spacing. Since the theoretical resolving ability of WETST is at the scale of λL (L is the traveling distance) , the grid spacing of the inverse grid should be even larger. Meanwhile, the resolution of the inversion results also relies on the coverage of seismic data. Checkerboard resolution tests are good measures on the resolving ability of an inverse algorithm with chosen seismic data and model parameterization. Therefore, checkerboard resolution tests are conducted to verify the chosen data, inversion grid, and the WETST algorithm in Sect. 3. However, due to the demanding computational cost, we do not test the limit of the inversion grid size in this study.

After the calculation of sensitivity kernels and the interpolation of relative velocity perturbation δc(x)/c(x) on inversion grid, tomographic Eq. () could be discretely expressed as a linear system b=AX, where b=[bm]M×1 is the travel-time residual vector (bm=Tmobs-Tmsyn and m is the index for a particular travel-time record), A=[am,n]M×N is the Fréchet matrix, and X=[Xn]N×1 is the unknown velocity perturbation vector. Usually, the limited data coverage deems this inversion an ill-posed problem, and b=AX is solved instead by minimizing the following regularized objective function χ(X)=12(AX-b)TCd-1(AX-b)+ϵ22XTCm-1X+η22XTDTDX, where D is a first derivative smoothing operator, and ϵ and η are the damping parameter and the smoothing parameter, respectively e.g.,. Cd and Cm are the a prior data and model covariance matrices and reflect the uncertainties in the data and the initial model. Here we assume that both Cd and Cm are identity matrices.

Either the LSQR solver or non-linear conjugate-gradient method can be used to solve the optimization problem (Eq. 7) as discussed in paper I . We choose to use the LSQR solver in this study. The solution of the minimization problem () can be obtained by solving the equivalent linear system using the LSQR solver AϵIηDX=b00. The choice of the damping and smoothing parameters involves some degree of subjectivity. Analysis of the trade-off between the data variance reduction and the model smoothness may help the selection of optimal damping and smoothing parameters . After the Vp and Vs models are updated, the Poisson's ratio (σ) image can be determined based on the relation δσσ=2Vp2Vs2Vp2-2Vs2Vp2-Vs2δVpVp-δVsVs, which is derived from the relation between Poisson's ratio and Vp/Vs ratio VpVs=2(1-σ)1-2σ. Clearly, the reliability of the Poisson's ratio result depends on the accuracy of both recovered Vp and Vs structures.

In order to obtain the discrete velocity perturbation X in Eq. () at each iteration, the damping parameter ϵ and the smoothing parameter η should be determined beforehand. In practice, these two parameters can be chosen via a trade-off analysis of data variance σd2 and model variance σm2 . For the sake of computational efficiency, the unbiased data variance is approximated by σd2≈1M-1∑i=1MTiobs-Tisyn-∑j=1NaijXj-d¯2, where the data average d¯ is estimated as d¯=1M∑i=1MTiobs-Tisyn-∑j=1NaijXj. The unbiased model variance σm2 is calculated using the formula σm2=1N-1∑i=1N(Xi-X¯)2, with X¯=∑i=1NXi/N is the mean of X. The trade-off analysis tries to find optimal damping and smoothing parameters that reduce most of the data variance without giving rise to too large model variance . For the checkerboard resolution tests, we search the damping parameter ϵ in the range [0.1,2.0] with a step of 0.02, but set the smoothing parameter as η=0 at each iteration to reflect the knowledge that the inverted structures are not smooth and have perturbations of opposite signs at neighboring nodes. Figure  shows the trade-off curves for both P wave and S wave checkerboard resolution tests at the first three iterations. Based on the L curve method e.g.,, we choose the optimal damping parameter ϵ for P wave or S wave test at each iteration near the corner of the corresponding trade-off curve. For example, to obtain the P wave velocity model m2 from the starting model m1, the optimal damping parameter in Eq. () is chosen as ϵ=0.42 which gives the data variance σd2=1.571×10-4 s2 and model variance σm2=11.59×10-4 (Fig. a). Note that the model variance is calculated with respect to the model in the previous iteration. Since the data variance is significantly reduced from model m2 to m4 and the value of the data variance in model m4 is very small for either P wave or S wave checkerboard test (Fig. ), we stop the iteration procedure at the fourth model m4.

Figures  and  show the iterative results of checkerboard tests at five representative layers in the crust for the P wave velocity (Vp) and S wave velocity (Vs) structures, respectively. Generally speaking, the checkerboard patterns are well resolved by WETST in the source area of the Landers earthquake. This indicates that both P wave and S wave data coverages are adequate enough, and the tomographic results inverted based on these data are reliable and can be used for further interpretation. More specifically, the checkerboard patterns at the five layers are almost recovered even at the first iteration (Fig. a–e and a–e), and the subsequent iterations only slightly refine the models (Fig. f–o and f–o). For both P wave and S wave tests, WETST has higher resolution in the upper crust (0–5.5 km) and middle crust (5.5–16.0 km) than that in the lower crust (>16.0 km). This may be due to two main reasons. First, as most of the 2041 earthquakes used in this study are located above 20.0 km (Fig. a), the inversion grid nodes in the upper and middle crust are sampled by more data than those in the lower crust, which provides better constraints to the anomalies in the upper and middle crust. Secondly, the inversion grid nodes in the lower crust are mainly covered by travel-time sensitivity kernels for first arrivals at long epicentral distances as shown in Fig. . The resolving ability of the travel-time data is proportional to the width of the first Fresnel zone proportional to λL, where λ is the wavelength and L is the traveling distance e.g.,. A long traveling distance would result in relatively low resolution. In addition, the edges of the model range are likely to have poor resolution due to the lack of well crisscrossed kernels therein.

To further investigate the recovery ability of our tomographic method WETST, we calculate the structural similarity (SSIM) index ζ between the inverted model and the input checkerboard model . The SSIM index ζ between two velocity (or other positive physical parameter) models A and B is defined as ζ(A,B)=2μAμBσABμA2+μB2σA2+σB2+0.5, where μA, μB, σA, σB, and σAB are the average of A, average of B, variance of A, variance of B, and covariance of A and B, respectively. The 0.5 is added to ensure that SSIM index ζ is in the range [0.0,1.0], and it is 1.0 only when A and B are identical . Table  shows the SSIM indices between the iteratively recovered P wave and S wave velocity models (m2 to m4, Figs.  and ) and the input checkerboard models at seven vertical layers. It can be observed that the SSIM indices at depths less than 21.0 km generally approach 1.0 through the iterations for both P wave and S wave tests. The recovery rates of the final P wave velocity and S wave velocity models m4 are above 0.9 at the depths less than 21.0 km and greater than 0.78 at the depth of 21.0 km, again indicating that the heterogeneities from the surface to the depth of 21.0 km can be well resolved in this study. However, the SSIM indices at the depths 28 and 40 km are only around 0.5, implying decreased resolution in the lowermost crust and the uppermost mantle. It is worth noting that the SSIM indices at 1.0 km are smaller than those at 5.0, 10.0, and 15.0 km, which is probably caused by the better crisscrossing of the travel-time sensitivity kernels of earthquakes below 3.0 km at 5.0 km, 10.0 km, and 15.0 km depths than that at 1.0 km depth (Figs.  and ). The generally well-recovered Vp and Vs structures in our checkerboard resolution tests also imply that reliable Poisson's ratio structures can be derived from this tomographic study. We have also conducted other checkerboard resolution tests for noise data, as summarized in the Supplement. All these resolution tests give us confidence that WETST should be able to generate high-resolution tomographic results for the source area of the Landers earthquake.

The optimal regularization parameters ϵ and η should be determined to update the tomographic models at each iteration, similar to those in the checkerboard resolution tests. In this case, we search the optimal damping parameter ϵ in the range [6,40] with an interval of 1 and the optimal smoothing parameter η over [2,100] at a step of 2. In the searching procedure, we first set the smoothing parameter η=0 and find the optimal damping parameter ϵ based on the L curve method. With the optimal damping parameter ϵ, we then determine the optimal smoothing parameter η in the searching region. For both P wave and S wave tomographic inversions, Fig.  shows the trade-off analysis of data variance σd2 and model variance σm2 along with different damping and smoothing parameters throughout the iterations. The optimal damping and smoothing parameters are also indicated in Fig. . After each model update, we compute the root mean square (rms) value of the travel-time residuals using the formula rms=1M∑i=1MTiobs-Ti,ksyn2, where Ti,ksyn is the arrival time of the i-th record in model mk. Table  shows the values of rms. For both P wave and S wave results, we can find that rms monochronically decreases from m1 to m4 .

Figure  further shows the distributions of P wave (Fig. a–c) and S wave (Fig. d–f) travel-time residuals Tobs-Tsyn in models m1-m4. It is clear that travel-time residuals gradually become more centered around 0.0 s over iterations, indicating an overall reduction in total travel-time misfit. Since there is no significant decrease in rms (Table ) from m3 to m4, we stop our iteration at the fourth model for both P wave and S wave inversions, and m4 is viewed as the final tomographic model used for interpretations in the following sections.

We present iteratively updated map views of Vp (Fig. ) and Vs (Fig. ) models at five representative depths for the Landers earthquake area. It can be observed that the general patterns of Vp and Vs revealed by models m2-m4 are almost the same, with only slight increase in the amplitudes of velocity anomalies over iterations (Figs.  and ). This is consistent with the significant rms reduction from m1 to m2, and minor reduction in the following updates, as shown in Table . However, it should be also noted that velocity anomalies near the boundaries of the study area become more clear over iterations, which agrees with the checkerboard resolution tests showing increased recovery from m2 to m4 (Table ), especially in the boundary regions (Figs.  and ). These results imply the necessity of iteratively improving the velocity models, even though the patterns of velocity anomalies could be almost recovered in the first iteration based on WETST with the LSQR solver.

We summarize the main features of the final tomographic model m4. Map views of Vp (Fig. k–o) and Vs (Fig. k–o) reveal large velocity variations of up to ±8 %, which indicate strong lateral heterogeneities in the model region. The epicentral areas of the Landers, Big Bear, and Joshua Tree earthquakes exhibit clear lateral velocity contrasts from the surface to about 15.0 km depth (Figs. k–n and k–n). In the Mojave block , north of the San Andreas fault, high Vp and Vs anomalies are generally visible at the shallow depth 1.0 km (Figs. k and k) and negative velocity perturbations exists in the middle crust (Figs. m, n and m, n). Similar depth variation of the velocity structures in this region was also reported by . In the upper crust, low-velocity anomalies (Figs. k, l and k, l) exist along the San Andreas fault (SAF) and the San Jacinto fault (SJF) but only beneath the northwestern portion of the Elsinore fault (EF) . Additionally, a significant high-velocity zone is visible between the SAF and the SJF, which results in strong velocity contrasts across the two faults near the surface . The high-velocity zone between the EF and the north portion of the SJF may indicate a reversal in the velocity contrast polarity along the SJF at around 5.0 km depth (Figs. l and l) . In the middle crust, the SAF, the SJF, and the EF roughly show relatively high-velocity anomalies (Figs. m, n and m, n) . However, in the lower crust (21.0 km), low-velocity anomalies are generally reported along these fault systems (Figs. o and o). We will discuss this low-velocity feature in detail in the next section. Beneath the Salton Trough (ST), which is a sediment-filled graben near the southern part of the SAF , a pronounced low Vp and Vs anomaly exists in the upper crust (Figs. k, l and k, l), and high P wave velocity structures are revealed in the middle and lower crust (Fig. m–o). This is consistent with the results of .

A series of vertical cross-sectional views from the surface to 40 km depth for Vp, Vs, and Poisson's ratio σ structures are shown in Figs.  and . Since both Vp and Vs structures are almost well recovered in the crust (Figs.  and and Table ), Poisson's ratio models in Figs. c, f, i and c, f, i can be viewed as being reliably determined based on Eq. (). Note that the map views of the iteratively updated Poisson's ratio σ structure are included in the Supplement (Figs. S3–S5). Figure  shows three cross sections along the profiles through the hypocenters of the Landers earthquake, the Joshua Tree earthquake, the Big Bear earthquake, and the 1999 Hector Mine earthquake where profile AB is nearly parallel to the fault zone of the Landers earthquake. It can be observed that the Landers mainshock is located in a high-velocity, low Poisson's ratio anomaly (Fig. a–c and g–i). Additionally, the hypocenters of the Joshua Tree, Big Bear, and 1999 Hector Mine earthquakes are at or near high-velocity and low Poisson's ratio anomalies (Fig. ). By inverting P wave arrival times from aftershocks of 1992 southern California earthquakes, also reported that high Vp anomalies occur at or near nucleation sites of the Joshua Tree, Landers, and Big Bear mainshocks. Both velocity and Poisson's ratio structures change drastically around the source areas of the Landers mainshock and the other three large earthquakes. Material properties of the source areas of the four large earthquakes are consistent with those of the brittle seismogenic layer, which is characterized by high velocity and low Poisson's ratio . A prominent feature of the vertical cross sections along profiles CD and EF is the low Vp, low Vs, and high Poisson's ratio structure to the west of the Big Bear mainshock hypocenter in the lower crust (Fig. d–i), which has been interpreted as a ductile and weak region . A low-velocity and high Poisson's ratio structure is also visible in the lower crust close to the hypocenter of the Landers mainshock (Fig. a–c and g–i). In addition, tomographic results and seismicity along the profile AB confirm the conclusion of that shallow earthquakes mostly occurred in high Vp regions and mid-crustal earthquakes occurred in low Vp zones. Seismicity along profile AB is mainly the aftershocks of the Landers earthquake (Fig. a–c), and it can be observed that to the south of the Landers mainshock hypocenter, seismicity strikes across the Joshua Tree aftershock zone, extends about 40.0 km south of the epicenter of the Landers mainshock, and terminates within a few kilometers of the SAF. Immediately following that of the Salton Trough, a low-velocity and high Poisson's ratio anomaly exists near the surface . Additionally, to the north of the Landers mainshock, aftershocks extend about 60.0 km to the Camp Rock fault and are surrounded by low-velocity and high Poisson's ratio rocks beneath them .

Figure  shows the vertical cross sections along the Elsinore fault (EF), the San Jacinto fault (SJF), and the San Andreas fault (SAF). Beneath the northwest segment of the EF, a low-velocity and high Poisson's ratio anomaly is visible in the upper and middle crust, underlaid by a high-velocity and low Poisson's ratio structure (Fig. a–c). These are contrary to the structural properties under the central and southeast sections of the EF, which generally exhibit high velocity and low Poisson's ratio in the upper and middle crust and low-velocity and high Poisson's ratio beneath (Fig. a–c). Seismicity along the EF is generally focused between 5.0 km and 15.0 km. Velocity and Poisson's ratio models reveal complex patterns beneath both the SJF and the SAF (Fig. d–i). Alternating high- and low-velocity variations can be observed along the faults near the surface, which can be interpreted as manifestations of the complex surface geological patterns. However, along both the SJF and the SAF, we can observe a layer with low velocity and high Poisson's ratio at the depth of about 5.0 km. Right beneath this layer, high-velocity and low Poisson's ratio structures exist in the middle crust. Seismicity along the SJF and the SAF mainly occurred in this high-velocity and low Poisson's ratio region. Additionally, the seismicity along the SJF is much more active than that along the SAF for study area . The lower crust is generally dominated by low-velocity and high Poisson's ratio structures. Specifically, near the southeast sections of the SJF and the SAF which are close to the Salton Trough, there are mainly low-velocity and high Poisson's ratio structures at shallow depths and high-velocity and low Poisson's ratio anomalies in the middle and lower crust. These features are consistent with the extension and crustal thinning of the Salton Trough region .