Seismic anisotropy accrued by seven unusually deep local earthquakes (between 50 and 60 km) in the Albertine Rift: implications of asthenospheric melt upwelling

We investigated the primary mechanisms triggering the S-wave splitting (SWS) of seven unusually deep local earthquakes (between 50 and 60 km) which originated in the lithosphere beneath the Rwenzori region. We attempted to develop an understanding of the relationship between anisotropic structures in the lithosphere and tectonic deformation processes. A total of 12 out of 44 waveforms showed evidence of SWS on their polarization diagrams. The fast-wave direction (φ) and delay-time (δt) were estimated using the covariance matrix and the cross-correlation coefficient methods, respectively. We observed a clockwise rotation of φ-directions (NW - SE and ~ENE - WSW) at stations located in the rift valley. We related this pattern of φ-directions to anisotropic fabric, probably lattice-preferred orientation of preexisting olivine, whose a-axes are aligned with ESE absolute plate motion (APM) vector. At stations located outside the rift valley, however, we observed WNW - ESE and NNW - SSE patterns of φ-directions. We associated these patterns to the shape-preferred orientation of structures frozen in the lithosphere that are aligned with the present-day APM direction. We observed δt values ranging between 0.04 ± 0.01s and 0.43 ± 0.02 s, which decrease with distance away from the rift axis. This further supported our concept that the anisotropy observed at stations located on the moving plate is related to aligned melt inclusions frozen in the surrounding lithosphere. We further observed that the δt values increase linearly with ray-path length, which could indicate a fairly uniform anisotropy between 50-km and 60-km depth. Our study reported no evidence of multi-layer anisotropy beneath the Rwenzori region.


Introduction
Studies of seismic anisotropy in the mantle have led to advancements in constraining upper mantle deformation by resolving directional features, which may provide information about structures in the upper mantle (Gao et al. 2010). Understanding the structure of the upper mantle, the way it deforms and flows, has been attributed to the quantity and quality of seismic data that have been recorded during seismic studies conducted around the world. One common indicator of seismic anisotropy is the strain-induced lattice-preferred orientation (LPO) of olivine crystals, whereby a-axes of olivine align parallel with the direction of maximum shear stress (e.g., Liu et al. 2008). Mantle melting has also been observed to influence seismic anisotropy in magmaproducing regions (e.g., Gashawbeza et al. 2004). Alignment of melt-filled pockets can induce seismic anisotropy, which does not depend on LPO (e.g., Long and van der Hilst 2006). Walker et al. (2004), in their studies across the NW -SE striking Anza Graben in Kenya reported the presence of channels in the base of the lithosphere. They argued that these channels could have provided guides to plume-related flows. Walker et al. (2004) further observed that partially molten inclusions like dikes can penetrate the lithosphere if there is a steady-state active source of partial melt in the asthenospheric mantle. These dikes can trigger S-wave splitting (SWS) due to the large velocity contrast between magma and surrounding rocks resulting in considerable delay-times (δt). Bagley and Nyblade (2013) similarly observed that seismic anisotropy in East Africa is strongly influenced by magmatic flow in the lower mantle. Their studies have provided a link between the processes in the lower mantle and the tectonic deformation of the Earth's surface. In the Western Rift of the East African Rift System (EARS), Tepp et al. (2018) observed rift-parallel fast-wave directions (φ). They argued that the spatial variations of the φ-directions could imply that the nature of mantle flow is a result of the deflection caused by the presence of deeply rooted cratons. Elsewhere in the EARS, Ayele et al. (2004) observed that the anisotropy in the northern Ethiopian Rift is associated with the alignment of the a-axes of olivine in the asthenosphere with the ridge axis. They concluded that this could be caused by magmatic material flowing laterally to fill the gap that was developed during the lithospheric extension.
In this paper, we present the results of research conducted on a cluster of seven unusually deep local earthquakes, which originated at extraordinary depths (between 50 km and 60 km) beneath the Rwenzori region. These earthquakes were recorded between 2nd and 16th September 2006 by 12 seismic stations, which were set up in the Rwenzori region of the Albertine Rift during the RiftLink research project (see Lindenfeld and Rümpker 2011). The Albertine Rift, a roughly NE to ENE trending sector of the EARS, extends along the western boundary of Uganda and the Democratic Republic of Congo. Within the Albertine Rift are the 5000m-high Rwenzori Block Mountains. Koehn et al. (2008) reported that the Rwenzori Block Mountains were captured by two rift segments: the Albertine Rift segment, west of the mountains extending from Lake Edward in the south to Lake Albert in the north, and the Lake George Rift segment east of the mountains extending from Lake Edward terminating north of Lake George (shown by gray dotted arrows in Fig. 3b). In this study, we focused attention on investigating the primary mechanism triggering the SWS of these deep local earthquakes. We attempt to develop an understanding of the relationship between anisotropic structures in the upper mantle and tectonic deformation processes beneath the Rwenzori region.
2 Accuracy of hypocenters for the seven earthquakes Lindenfeld and Rümpker (2011) located these seven deep local earthquakes using the HYPOCEN-TER location algorithm of Lienert and Havskov (1995). To ascertain the reliability of the location algorithm, Lindenfeld and Rümpker (2011) relocated the hypocenters using both the IASP91 velocity model of Kennett and Engdahl (1991) and the double-difference algorithm (hypoDD) of Waldhauser and Ellsworth (2000). They concluded that in all cases, there were no significant deviations in the depth distributions and that the event locations were reliable with a precision better than 5 km. Therefore, the depths observed for these seven deep local events are robust and cannot be related to artifacts caused by cycle-skipping on the waveforms. Wölbern et al. (2010) used observations estimated using teleseismic receiver functions to constrain the thickness of the crust beneath the Rwenzori region. Using the method of Zhu and Kanamori (2000), they estimated a crustal thickness of about 30 km on the eastern rift flank, 20 km and 28 km beneath the northern and central parts of the Rwenzori block, respectively. It is therefore unlikely for these deep local earthquakes to have originated from the crust, but rather from the lithospheric mantle beneath the Rwenzori region. It was initially not clear how these deep local earthquakes could have originated from such an unexpected depth. Schmeling and Wallner (2012) associated these deep-focus earthquakes to the riftinduced delamination process beneath the Rwenzori block. However, their findings were inconclusive. Lindenfeld and Rümpker (2011) therefore concluded that these deep earthquakes did not originate directly below the Rwenzori block and that magmatic impregnation of the mantle lithosphere described by Foley (2008) could be a likely cause of seismic radiation at such an unexpected depth.

Data and methods
We used data that were recorded by EDL and REFTEK data loggers, coupled with a Güralp CMG-3T and Mark L-4C3D seismometers, respectively. These data loggers were configured to operate in continuous and trigger modes, respectively, and were sampling at 100 samples/second. We applied the short-term average (STA) and long-term average (LTA) to the continuous data streams to detect and extract local events. We used an STA and LTA of 3 s and 30 s, respectively, for the time windows and an STA/LTA ratio of 3.0 as a trigger to an event. Figure 1 shows the seismic waveforms of seismic event 06902021330 that was recorded at several recording stations. We adopted a critical angle for the S-wave window of ≤ 34°, which according to Booth and Crampin (1985) is adequate for minimizing contaminations from S-to P-wave mode conversions near the surface. Booth and Crampin (1985) further observed that these contaminations distort the amplitude and phase of the recorded wave yielding a nonlinear particle motion. These seven deep local earthquakes yielded 44 waveforms for SWS analyses ( Table 1). The horizontal waveforms were initially Butterworth filtered between frequencies 0.1 and 10 Hz to separate the dominant frequency from noise. We applied both the automatic and visual display techniques to the horizontal channels to estimate the SWS parameters. Each of these techniques however has its advantages and drawbacks. The most optimal approach is to combine the advantages of both displays (polarization diagrams and rotated seismograms) and automatic (cross-correlation and covariance) techniques for measuring the splitting parameters (Gao et al. 2006). We applied the particle motion analysis of Crampin et al. (1985) to estimate the SWS parameters. Only those traces, which showed abrupt changes in particle motion on their polarization diagrams (see Fig. 2b) were selected for further analysis. Consequently, only 12 out of the 44 waveforms showed evidence of abrupt changes in particle motion (see Table 1). More detail on the SWS analysis is given in the supplementary information (Figs. S1 to S11). The rest of the traces were rejected due to either low signalto-noise ratio and/or lack of anisotropic evidence (events with dashes in Table 1).
Because we analyzed horizontal seismograms for SWS, we only applied the horizontal projection of φdirection. Therefore, the waveforms that showed a has been plotted over the N-S component seismogram (black waveform). These seismograms have been plotted at their respective epicenter distances. The time shown is that after the event time which is 2006-09-02 02:13:30. The P-and S-wave arrivals have been marked by the vertical red lines Table 1 Showing 44 waveforms recorded at the 12 seismic stations. The data logger/seismometers are given as E/G, E/M, and R/M corresponding to EDL PR6-24/Guralp GMG-3T, EDL PR6-24/Mark L-4C3D, and REFTEK 72A-07/Mark L-4C3D, respectively. The dashes imply null measurements due to either low signal-to-noise ratios and/or lack of anisotropic evidence beneath the seismic station (nulls). The coordinates slon and slat are seismic station longitudes and latitudes respectively. maxXcorr represents the maximum cross-correlation coefficients and their corresponding lags. cruciform particle motion were rotated into the fast and slow coordinate system (Fig. 2c). For each measurement, we selected a time window, which included the earliest S-wave arrival, and the φ-direction in the horizontal plane was estimated using the covariance method. The principal application of the covariance method was initially based on the theory of polarization filters. Shimshoni and Smith (1964) used this method to constrain the rectilinearity of the recorded seismograms. Since then, several researchers have used this theory to develop applications that detect phase arrivals on seismograms (e.g., Montalbetti and Kanasewich 1970). The covariance matrix method uses the covariance matrix (CM) of the horizontal channels (N, E) for "n" set of points taken from within a specific time window to estimate the φ-direction.
We diagnosed the covariance matrix and estimated the φ-direction considering the eigenvector associated with the largest eigenvalue for the coordinate directions N and E. Given that λ 1 and λ 2 are the largest and next largest eigenvalues of covariance matrix, respectively, the function f (λ 1 , λ 2 ) will tend to unity when rectilinearity is high (λ 1 » λ 2 ) or it will tend to zero when rectilinearity is low (i.e., when the two principle axes approach one another in magnitude).
The polarized fast and slow traces will split in time when they come across an anisotropic media, and consequently, the two S-waves will appear as identical wavelets separated by a certain δt. In an isotropic environment, however, the splitting is preserved along the ray path (Díaz et al. 2006). We therefore estimated the δt Fig. 2 An example of SWS of a seismic waveform. a Original waveforms in the horizontal coordinate system b cruciform particle motion showing evidence of SWS, and c the rotated waveforms (into fast and slow horizontal channels). The time window (vertical grey shading) shows the difference between the onsets of the fast and slow shear-waves and will be used to estimate the δt. d Results of the normalized cross-correlation coefficients plotted against lag (the lag corresponding to the maximum cross-correlation coefficient shown in the brackets was used to estimate the δt) between the fast and slow horizontal channels using the cross-correlation coefficient method.
This method was initially implemented by Fukao (1984). Since then, the cross-correlation technique has been successfully applied in most automatic measurement techniques to estimate the δt both above small earthquakes in the crust and upper mantle anisotropy. This technique can be used to estimate the δt from lags of cross-correlations of the preferentially rotated wavetrains. The SWS signals tend to be noisy as the S-waves are disturbed by the P-wave coda and the slower split Swaves are disturbed by the coda of the faster split Swave. Studies by Crampin and Gao et al. (2006) have shown that it is often challenging to choose suitable ends to windows for cross-correlation. An ideal time window should begin before the fast S-wave arrival and end after the slow S-wave arrives, but before the scattered coda waves appear. In our study, we selected a window that included the earliest S-wave arrivals on both horizontal channels such that the fast and slow Swaves are almost orthogonally polarized and that the influence of the P-wave and S-wave codas is minimal. We then visually examined each record on a polarization diagram (Fig. 2b) for quality control. However, using the cross-correlation function to measure the similarity between the two seismic traces is usually misleading especially if there are large differences in the energies of these two traces. Therefore, we scaled the cross-correlation function by a factor that is determined by the energy levels of the analyzed seismic traces. We then estimated the measure in the similarity between the two seismic traces (Fig. 2d) using the normalized crosscorrelation function (NormXcorr). Splitting analyses of all the other waveforms are provided in the supporting information.
It is always essential to evaluate the quality of φ and δt to ensure the reliability of the splitting measurements. We therefore adopted the standard error of the mean technique, which is a standard statistical approach to access the uncertainties in the splitting parameters at a 95% confidence level. We further explored the robustness of the SWS measurements using the approach adopted by Gao et al. (2006). We introduce three grades of quality: good (g), acceptable (a), and unacceptable (u). We define two values SNR g and SNR a specifying good and acceptable signal-to-noise ratios (SNR), respectively. If SNR ≥ SNR g , then the quality of the polarization is set to good. If SNR a ≤ SNR ≤ SNR g , then the quality of the polarization was set to acceptable; otherwise, it was set to unacceptable. We used the program to evaluate noise and signal to noise ratio of Sharma et al. (2016) to estimate the SNR. We similarly selected SNR a to be equal to 3.0 and SNR g equal to 5.0.

Results
We present our SWS measurements as thick black lines that we scaled by δt and aligned by φ (Fig. 3a). The red cones shown in the background of each SWS measurement represent the uncertainties in the splitting results at a 95% confidence level. To avoid the cluttering of SWS measurements at a station, we plotted the measurements in the vicinity of the seismic station at arbitrary locations, specifically those where more than one measurement was analyzed. We plotted the earthquake in the middle of each measurement in order to associate it with its corresponding SWS measurement. We partitioned the study area into two sub-regions, i.e., that in the rift valley ( Fig. 3b: red ellipse) and that on the ESE propagating Victoria microplate (Fig. 3b: blue ellipse). Studies conducted by Calais et al. (2006) have shown that the Victoria microplate is part of the old Congo-Tanzanian Craton which according to Koehn et al. (2010) is still connected to the northern tip of the Rwenzori Mountains, an extraordinary basement block mountain that developed within the western branch of the EARS. Figure 3a was quite revealing in several ways. First, we observed a clockwise rotation of φ-directions from NW -SE and~ENE -WSW recorded at stations in the rift valley (BUMA, KABA, and NTAN). Second, at stations that were located on the Victoria plate (BUTU, KASS, and MIRA), we observed WNW -ESE and NNW -SSE orientation patterns of φ-directions which align with the ESE APM vector.
We further observed that our φ-directions are consistent with the asthenospheric mantle flow model of Sleep et al. (2002). Table 1 shows the station-earthquake pairs with δt values ranging between 0.04 ± 0.02 s and 0.43 ± 0.02 s. Figure 4a shows that the magnitude of splitting measurements varies across the rift. The largest δt values were observed along the edges of the faultbounded rift valley, which decay away from the NNEtrending rift axis (also see Fig. 3a). We further analyzed the anisotropy at stations that registered the same event (Fig. 5). In this case, we used event 060902021330 since this event was recorded at a number of recording stations. We however observed similar patterns as shown in Fig. 4.
In this study, almost all the seismic stations that recorded the deep earthquakes registered at least a null result except at BUMA and BUTU, probably because few SWS observations at these stations were made (Fig.  3b). In Fig. 3b, we have plotted the initial polarization directions (black bars) to represent the null measurements and the gray bars to represent the direction perpendicular to the initial polarization direction. Null measurements in a seismic dataset occur (1) if there is an absence of anisotropy beneath the seismic station, (2) if the initial polarization coincides with either the fast or the slow azimuth of the fast axis and (3) due to vertical anisotropy with an a-axis of the olivine crystal aligned in the direction of S-wave propagation. In all cases, the Swave will not split (Savage 1999). Studies by Silver and Chan (1991), Barruol et al. (1997), andFouch et al. (2000) have shown that null measurements are often treated independently or even neglected in SWS studies. In total, 32 null measurements were observed. We neglected these null measurements in our discussion since no splitting was identified, probably because of an absence of seismic anisotropy beneath the recording station. Savage et al. (1996) argued that when nulls are measured at a wide variety of azimuths for S-waves with steep incident angle, this implies that the medium is either effectively isotropic or transversely isotropic with a vertical symmetry axis. We have provided an example of the analysis of a null measurement in the supplementary information (Fig. S12). Issues regarding the relationship between strain and anisotropy have been a controversial and muchdisputed subject in the field of seismology (e.g., Savage 1999). To commence our discussion, we initially examined the kinematics and magnitude of active and past deformation mechanisms present in the lithosphere beneath the Rwenzori region. We attempted to associate these mechanisms with the observed anisotropy in the region. It seems likely that both aligned melt intrusion zones in the lithosphere and aligned asthenospheric flow mechanisms are at play in study region. Factors which are thought to be influencing seismic anisotropy in the mantle have been explored in several studies. For example, Karato and Jung (2003) have associated seismic anisotropy in the mantle to the alignment of the a-axis of olivine through dislocation creep. Numerical models of Tommasi et al. (1999) have shown that LPO of pre-existing olivine tends to be reactivated in a direction that combines both strike-slip motion and extension regimes, parallel and orthogonal to the rift axis, respectively. They argued that this deformation regime would consequently invoke a composite LPO of preexisting a-axis of olivine, oblique to the orientation of the rift axis. Similar inferences have been made to support this oblique deformation by Rümpker et al. (2003). Numerous studies have reported the nature of olivine. For example, Walker et al. (2004) reported that olivine comprises a significant fraction of the upper mantle. They observed that when olivine is deformed via dislocation creep, one or more of the three olivine crystallographic axes develop LPO, an indication of anisotropy. Gao et al. (1997) argued that the φdirections in the asthenospheric mantle are related to LPO of olivine's a-axis that is aligned with the present-day APM vector which according to Vinnik et al. (1992) can be related to asthenospheric shear. Gripp and Gordon (2002) suggested a 285 o ± 45 o APM of the African plate at 15 ± 3 mm/yr. Stamps et al. (2008Stamps et al. ( , 2020 have reported that the small Victoria microplate, located east of the Rwenzori block, is propagating in an ESE direction at a velocity of 2.1 mm/year relative to the Nubian plate that is located west of the Rwenzori block. An ESE propagating direction of the Victoria microplate at a velocity of about 5 mm/year has been reported by Calais et al. (2006). Interestingly, our results of φ-directions at stations located in the rift valley are consistent with the present-day rift-perpendicular ESE APM vector, which is not surprising given the long history of extension in the region.
In this study, we sought to address two questions: (1) is it possible that dynamic stretching of the lithosphere beneath the Rwenzori region occurred, which could have resulted into a lithospheric thickness of < 50 km especially around the rift axis, and that the lithospheric thinning was compensated for by balancing the ascent of asthenospheric melt along the inverted valleys that follow the convex-like base of the lithosphere? (2) If the lithospheric thinning was not sufficient enough, is it possible that the asthenospheric melt extruded into a thick lithosphere to shallow depths of < 50 km? If any of these hypotheses is correct and considering the location of these deep earthquakes (which are located outside the rift valley; see Fig. 3a), then it is likely that the travel-path of the earthquakes to the stations located in the rift valley (BUMA, KABA, and NTAN) could have traversed through the upwelled asthenospheric melt (Fig. 6) triggering olivine-induced SWS assuming LPO of olivine.
Several studies have estimated the thickness of the lithosphere and also reported the presence of magmatic intrusions beneath the Rwenzori region. For example, Link et al. (2010) conducted geochemical analysis on volcanic rocks and reported lithospheric thinning ranging from > 140 km beneath Toro-Ankole (Western Uganda) to about 80 km beneath the Virunga volcanic field in the Democratic Republic of Congo. Wölbern et al. (2012) have placed the thickness of the crust at 30 km, the upper boundary of the lithosphere at 60 km, and the lithosphere-asthenosphere boundary at 120 km. There have been several reports which confirmed the presence of magmatic material beneath the Rwenzori region. For example, numerical models of Wallner and Schmeling (2010) have predicted the upwelling of a broad region of asthenospheric melt into the lithosphere below the Rwenzori block. Batte and Rümpker (2019) observed the occurrence of broad magmatic chambers beneath the Rwenzori region. Studies of Lindenfeld et al. (2012); Wölbern et al. (2010Wölbern et al. ( , 2012 similarly offered evidence that suggest the presence of magmatic intrusions into the lithosphere beneath the Rwenzori region. Ochmann et al. (2007) observed temperature anomalies that they associated with a hot degassing magmatic intrusion. They concluded that this magmatic intrusion could be a possible heat source for the nearby Sempaya Hotspring (red star in Fig. 3b). If all these concepts are correct, then it is likely that an upward penetration of asthenospheric melt into the lithosphere and the horizontal flow of asthenospheric melt at the top of the plume associated with rifting could be the most plausible cause of the observed φ-directions at stations BUMA, NTAN, and KABA. The magmatic material formed from decompression melting of the Fig. 6 Showing melt upwelling in the mantle lithosphere beneath the Rwenzori region. The locations of the deep events are also illustrated. The depths to the Moho and of upper mantle discontinuity have been reported by Wölbern et al. (2010Wölbern et al. ( , 2012. The SKS φ-directions are illustrated by the red-headed double arrows. The deep events φ-directions are illustrated by the blue double-headed arrows asthenospheric mantle could have intruded the lithosphere. This magmatic material could have been deflected by the moving plate resulting in the formation of anisotropic fabrics that are aligned with the APM vector. Studies by Meissner et al. (2006) have shown that magmatic melts from the mantle extruded into the lower crust causing heating and layering. These authors argued that ductile deformation is responsible for the alignment of anisotropic minerals leading to transverse isotropy or azimuthal anisotropy. Similar inferences have been made across the East African Plateau by Walker et al. (2004). It is therefore likely that thẽ ENE -WSW patterns of φ-directions observed at BUMA, KABA, and NTAN could be related to the alignment of the anisotropic fabric associated with the transcurrent flow of the upwelled asthenospheric melt. This melt was probably driven by plate movement, possibly during a prior event of rifting. Of course, we cannot rule out the fact that the measurements observed at some stations in the rift show significant scatter (e.g., KABA). These φ-directions, based on the assumption of a single anisotropic layer, revealed a conspicuous change in the φ-direction pattern from~ENE -WSW to NW -SE direction. This could probably be related to local structural irregularities or a more complex anisotropic symmetry beneath the station.
We continue to discuss the anisotropy observed at stations located outside the rift valley, i.e., on the ESE propagating Victoria microplate (KASS, BUTU, and MIRA). We could similarly argue that the observed φdirections at these stations are influenced by LPO of olivine's a-axes, which are aligned with the proposed ESE APM vector. However, since the lithosphere is rigid and given that asthenospheric upwelling of magmatic material in the lithosphere outside the rift valley is relatively small in amount, it is unlikely for this pattern of φ-directions to be associated with aligned a-axes of olivine as we presumed but rather with shape-preferred orientation (SPO) of structures that are trapped and frozen in the lithosphere and are aligned parallel to the present-day APM direction of the Victoria plate. Similar inferences have been made by Gao et al. (1997) during their study across the Kenya Rift. Factors that are thought to be influencing the SPO of water-filled or melt-filled dikes/lenses in the lithosphere have been reported in several studies. For example, Walker et al. (2004) observed that magma-filled lenses and/or partially molten dikes can induce effective anisotropy. Kendall et al. (2005) have reported increased splitting coupled with increased magma production near breakup in the Ethiopian Rift. They argued that the melt will solidify, drifting away from the mantle upwelling and consequently leaving a residual anisotropy predominantly due to crystal alignment. Studies by Foley (1992) observed highly potassic and silica under-saturated lavas beneath the Rwenzori region, which they related to small-scale veins within the lowermost lithosphere. This according to Foley et al. (2012) indicated a presence of small-scale dikes and melt-filled lenses below the crust between 60 and 140 km.
Previous research has shown that pockets/lenses of melt inclusions into the lithosphere can have a significant influence on observed seismic anisotropy. For example, Walker et al. (2004) observed the δt values to be large at the center around the rift axis but diminish with perpendicular distance away from the rift axis. They interpreted this as being due to the cooling of the dikes/magma lenses over time. A close inspection of our results revealed a significant decrease in the δt values with distance away from the rift axis (Fig. 4a). This observation further complements our previous conclusion that the observed anisotropy beneath the stations (BUTU, KASS, and MIRA) is due to rift structures, probably aligned melt inclusions frozen into the surrounding lithosphere. The NNW -SSE pattern of φdirection observed at station BUTU however deserves mention. This station was located at/close to the junction between the Rwenzori basement block and the Victoria plate. We speculate that the NNW -SSE pattern of φdirection observed at station BUTU could have been caused by the perturbation of the local stress field in the vicinity of the seismic station. Koehn et al. (2008) observed that the local stress field at the rift segment tips is perturbed as the Victoria micro-plate was being captured. They concluded that the perturbed stress field produces extensional structures that are almost aligned with the far field extension direction. Alternatively, it could have been a result of the clockwise rotation of the Rwenzori block (Koehn et al. 2008). These authors argued that this rotation of the block could have led to the formation of local compression related structures at corners where the micro-plates are still attached to larger-scale plates.
We cannot rule out the fact that a significant contribution of the observed δt is from crustal sources since the crust accounts for about 50 % of the travel path (20-30-km crustal thickness). Analyses of SWS on local earthquakes (focal depth ≤ 30 km) in the crust beneath the Rwenzori region have been reported by Batte et al. (2014). They observed a more-or-less rift-parallel pattern of φ-directions at stations located in the south of the study area and a highly heterogeneous pattern of φdirections at stations located north of the study area. The authors attempted to associate these observed φdirections with orientation patterns of fault surfaces measured on in situ rocks outcropping in the study area. They observed that orientation of the fault surfaces agrees with the observed φ-directions at some stations north of the study area. They therefore concluded that seismic anisotropy in the crust of the Rwenzori region could be partly controlled by alignment of fracture and/ or cracks. Batte et al. (2014) continued to associate the observed φ-directions with orientation patterns of foliation planes measured on exposed in situ rocks in the Rwenzori region. They reported a close correlation of φ-directions with the foliation fabric at some stations in the north of the study area which suggests manifestations of fossil strain probably associated with the last major orogenic episode. They therefore concluded that the observed φ-directions in the crust of the Rwenzori region are partly controlled by alignment of fractures [EDA hypothesis of Crampin 1991] and partly by seismic anisotropy associated with the Precambrian foliation and therefore bear the signature of the paleostress field. Batte et al. (2014) further reported an average δt of 0.04 s (amounting to~4 % anisotropy) although large δt values of up to 0.23s were observed NW of the Rwenzori Mountains. They associated these large δt values to the presence of stress aligned pockets of melt within the shallow crust. Homuth et al. (2016) conducted SWS analyses on SKS earthquakes recorded in the Rwenzori region. They reported rift-parallel φdirections that are consistent with horizontal transverse anisotropy, which is typical during the early stages of continental rifting. They concluded that these φdirections are related to rift-parallel intrusions or lenses in the lithospheric mantle located at depths between 60 and 120 km.
In our study, however, the φ-directions of the deep events were observed to align almost perpendicular to the φ-directions observed for SKS events. It is likely that the upward flow of the melt could have been deflected by the ESE propagating plate movement as it ascended into the lithosphere resulting into the E-W to ESE-WNW φ-directions observed for the deep events. Homuth et al. (2016) reported δt values ranging between 0.2 s and 2.0 s (average ≈1.0 s). In this study, we Fig. 7 A plot of average δt values against ray-path length. The figure reveals a general increase of δt with ray-path length which is indicative of a fairly uniform anisotropy between 50-km to 60km depth observed an average δt of~0.23 s (amounting to~23 % anisotropy), which was accrued by these deep local earthquakes. It is therefore clear that the major contribution of seismic anisotropy in the Rwenzori region was accrued by SKS earthquakes (~73 % anisotropy). Studies of, e.g., Shih et al. (1991) and Kaneshima and Silver (1995) have reported that variations in the splitting parameters for earthquake sources at different depths directly beneath a seismic station in the crust or mantle provides the best estimate of the variation of anisotropy with depth. It is however often difficult to determine the depth of the anisotropy that produces the observed signals. The relative estimates of the depth of anisotropy may be inferred from ray-path length-δt plots. In our study, we observed that the δt increases linearly with ray-path length (Fig. 7), which suggests a fairly uniform anisotropy between 50-km and 60-km depth. A similar inference has been made by Keir et al. (2005) in their study across the Ethiopian Rift.
We proceeded to investigate whether there is a possibility of a presence of multi-layer anisotropy beneath the study region. Studies by Silver and Savage (1994) have shown that the φ-direction and δt exhibit important properties of being only a weak function of incident angle and back azimuth for near vertical incidence (for teleseismic Swaves). We therefore attempted to examine the dependence of incoming polarization direction with the apparent φ and δt (Fig. 8). Contrary to expectations, we did not observe any evidence of presence of multi-layer anisotropy since there is no discernable dependency of the splitting parameters with incoming polarization direction. This could probable suggest a simple structure of a single anisotropic layer beneath the study region. It could also be due to the very few observations used and/ or to the poor coverage which strongly limited scope of investigation. Whatever the mechanisms that trigger SWS at such an unexpected depth, if our interpretations are correct, may suggest evidence of a profound deformation regime, which may have important implications on the understanding of the geotectonic evolution of the lithospheric mantle beneath Rwenzori region.

Conclusions
Studies of SWS were completed on seven unusually deep local earthquakes that originated between 50 and 60 km beneath the Rwenzori region. Only 44 waveforms were available for interpretation from which, 12 showed abrupt changes in the φdirections on their polarization diagrams. We observed WNW -ESE and EW orientation patterns of φ-directions at stations located in the rift valley (BUMA, KABA, and NTAN). We related these φdirections to the presence of anisotropic fabrics Fig. 8 Delay time (a) and fast axis azimuth (b) plotted as a function of earthquake's incoming polarization direction. No discernable pattern is observed to indicate an evidence of the presence of multi-layer anisotropy (probably LPO of olivine), that were developed as a result of the asthenospheric melt, which extruded into the lithosphere that was deflected by the ESE moving plate. These anisotropic fabrics are aligned with the APM vector. At stations located outside the rift valley (BUTU, KASS, and MIRA), however, we observed WNW -ESE and NNW -SSE patterns of φ-directions which we associated with the shape preferred orientation of melt-filled pockets trapped in the lithosphere that are aligned with the present-day APM direction. The δt values accrued by these deep events range between 0.04 ± 0.01 s and 0.43 ± 0.02 s. These δt values significantly decrease with distance away from the rift (MIRA being the farthest station from the rift valley). This further supports our concept that the observed anisotropy beneath the stations (BUTU, KASS, and MIRA) is due to rift structures, probably aligned melt inclusions frozen into the surrounding lithosphere. We further observed that the δt increases linearly with ray-path length, which could suggest a fairly uniform anisotropy between 50-km to 60-km depth. We did not observe any evidence of the presence of multi-layer anisotropy since there is no discernable dependency of the splitting parameters with incoming polarization direction.