Rift Focusing and Magmatism During Late‐Stage Rifting in Afar

Processes that facilitate the transition between continental rifting and sea‐floor spreading remain unclear. Variations in the spatial distribution of extension and magmatism through Afar and into the Red Sea are indicative of temporal evolution of the rifting process. We develop a time series of Sentinel‐1 interferometric synthetic aperture radar (InSAR) observations of ground deformation covering the whole Afar rift zone from 2014 to 2019, to study the distribution of extension. By incorporating Global Navigation Satellite System observations, we resolve 3D average velocities in the vertical, rift‐perpendicular, and rift‐parallel directions. Results show the spatial distribution of long‐wavelength deformation over the rift zone, as well as deformation at individual volcanic centers, including Dallol, Nabro, and Erta ’Ale. We find that in northern and central Afar, the majority of extension is accommodated close to the rift axis ( ± 15–30 km). In southern Afar, near the Nubia‐Arabia‐Somalia triple junction, amagmatic extension is distributed over 80–160 km, which may indicate an increase in rift focusing with rift maturity. We also observe rapid surface uplift and rift‐perpendicular extension at the Dabbahu‐Manda‐Hararo rift segment with velocities of 33 ± 4 mm/yr and 37 ± 4 mm/yr respectively. These are higher than the background extension rate of 18–20 mm/yr, but have decreased by 55%–70% since 2006–2010. The data suggests that this is due to an ongoing long‐lived response to the 2005–2010 rifting episode, with potential continued processes below the rift segment including a lower‐crustal viscous response and magma movement. Continued observations of surface deformation provide key constraints on tectono‐magmatic processes involved in rift development.

. The Afar rift zone with white triangles indicating Holocene volcanoes and key volcanoes highlighted in yellow. Simplified volcano-tectonic segments (VS) are shown in red: AFVS, Afdera; ALVS, Alayta; AGVS, Asal-Ghoubbet; DMHVS, Dabbahu-Manda-Hararo; EAVS, Erta 'Ale; MI, Manda-Inakir; NMER, Northern Main Ethiopian Rift; TAVS, Tat 'Ale (taken from Keir et al. [2011]). Simplified faults in central Afar and simplified rift margin faults are displayed as black lines (taken from Pagli et al. [2014]). A subset of Global Navigation Satellite System (GNSS) velocity vectors with 95% confidence error ellipses (blue arrows) from King et al. (2019) show the long-term plate motions. GNSS sites from King et al. (2019) (blue circles) are supplemented by additional fabricated GNSS sites with zero velocity on the stable Nubian plate (blue squares). Gray box outlines show the Sentinel-1 coverage from three ascending (T014A, T087A, and T116A) and two descending (T006D and T079D) tracks. Inset map shows the relative movement of the Arabian and Somalian plates to the Nubian plate, with sketched plate boundaries from Bird (2003). RSR, Red Sea Rift; GAR, Gulf of Aden Rift; MER, Main Ethiopian Rift. zones of rift-parallel diking, as seen for example at the Dabbahu-Manda-Hararo segment of the Afar rift zone (Wright et al., 2006); (c) a breakup phase where the lithosphere has thinned sufficiently such that new "oceanic" crust is effectively being formed at the ridge; they suggest the Asal rift in Djibouti (Ruegg & Pilger, 1975) is a good example of this mode. In this conceptual view, strain should become increasingly localized as magmatic continental rifting becomes increasingly "mature." However, Kogan et al. (2012) used geodetic measurements of extension across three segments of the rift system (Southern Main Ethiopian Rift, Central Main Ethiopian Rift, "Afar") to suggest that strain becomes more diffuse as magmatic continental rifting progresses.
In this paper we combine dense geodetic observations of crustal motion from Sentinel-1 interferometric synthetic aperture radar (InSAR) time series analysis with sparse point measurements from Global Navigation Satellite System (GNSS) data to produce spatially dense grids of 3D surface velocities for the entire Afar region for the period 2014-2019. We use these to map the degree of strain localization for magmatic and amagmatic rift segments and to test whether strain localizes as rifting progresses. The 2005-2010 rifting episode at the Dabbahu-Manda-Harraro segment (Grandin, Socquet, Doin, et al., 2010;Hamling et al., 2010) created a large deformation response in the lithosphere (Hamling et al., 2014;Nooner et al., 2009;Pagli et al., 2014); here we also examine how that transient feature has evolved through time. We also present the recent time history of deformation at eruptive centers in Afar.

Regional Setting
Upwelling of a mantle plume initiated rifting in Afar around 30 Ma with abundant flood-basalt volcanism, which has evolved into the ridge-ridge-ridge triple junction observed in Afar today (e.g., Furman et al., 2006;Hammond et al., 2013;Schilling, 1973;White & McKenzie, 1989;Wolfenden et al., 2004). Relative to the Nubian Plate, GNSS observations show that the Arabian Plate is moving at a rate of 18-20 mm/yr to the NE (ArRajehi et al., 2010;Doubre et al., 2017;McClusky et al., 2010;Viltres et al., 2020), accommodated by the opening of the Red Sea Rift; while the Somalian Plate is moving to the SE at  E 6 mm/yr (Birhanu et al., 2016;Saria et al., 2014), accommodated by the Main Ethiopian Rift (MER) which is the northern-most segment of the larger East African Rift.
The crust beneath the Afar rift zone is significantly thinned in comparison to the surrounding Ethiopian Highlands and MER. Crustal thicknesses range from 20 to 45 km in the Ethiopian Plateau, 18-30 km in central Afar, and 15-20 km in northern Afar Hammond et al., 2011;Lavayssière et al., 2018;Tiberi et al., 2005). Low seismic velocities indicate partial melt within the crust (Gallacher et al., 2016), particularly below rift segments in Afar (Hammond, 2014;Hammond & Kendall, 2016;Stork et al., 2013). Seismicity in the upper and lower crust along the Erta 'Ale volcano-tectonic segment also indicates the presence of melt below the Erta 'Ale and Alu-Dalafilla volcanic centers .
Active volcanism in Afar is largely concentrated within discrete rifting segments (e.g., Barberi & Varet, 1978;Hayward & Ebinger, 1996;Manighetti et al., 1998). The Erta 'Ale segment in the northern-most portion of the Afar rift zone is the immediate on-land expression of the Red Sea Rift . Erta 'Ale volcano on the Erta 'Ale segment is host to a lava lake with recent overflows in 2010 (Barnie, Oppenheimer, & Pagli, 2016;Field et al., 2012) and 2017, where a flank eruption indicated the presence of a shallow magma body at  E 1 km depth (Moore et al., 2019;Xu et al., 2017). At Gada 'Ale, magma withdrawal and normal faulting caused subsidence from 1993 to 1996 (Amelung et al., 2000), and a dyke intrusion fed from a magma chamber 2-3 km below Dallol was detected in 2004 (Nobile et al., 2012). The 2008 eruption at Alu-Dalafilla was sourced from a  E 1 km deep axis-aligned reservoir and a magma chamber at  E 4 km depth .
The Manda-Inakir and Asal-Ghoubbet volcano-tectonic segments in the southern portion of the Afar rift zone have also shown recent activity with the 1928-1929 eruption of Kammourta volcano in the Manda-Inakir volcano-tectonic segment (Audin et al., 1990), and the 1978 eruption at Ardoukoba volcano in the Asal-Ghoubbet segment (Allard et al., 1979;Tarantola et al., 1979). A post-rifting response was identified in the Asal-Ghoubbet segment to the 1978 eruption, with rift-perpendicular velocities decaying back to the long-term spreading rate 6-8 years after the eruption (Cattin et al., 2005;Ruegg & Kasser, 1987;Ruegg et al., 1993).

InSAR Velocity Methods and Applications in Afar
Methods for extracting a one-dimensional line-of-sight (LOS) displacement time series from a sequence or network of interferograms are well established. These small baseline algorithms utilize multiple interferogram connections between acquisition dates to produce a more robust estimate of the incremental LOS ground displacement than a simple stacking of interferograms (Berardino et al., 2002;Biggs et al., 2007;Lanari et al., 2007). This methodology may be automated by software packages such as Π E -RATE , and references therein), StaMPS (Hooper et al., 2012), GIAnT (Agram et al., 2013), and LiCSBAS  in order to obtain linear displacement rates and uncertainties at each pixel, while reducing the effect of common sources of error such as atmospheric and orbital delays. These methods may be supplemented by additional filtering to remove the atmospheric phase screen (APS) from the time series, by first high-pass filtering in time, then low-pass filtering in space to calculate the APS, which is then removed from the time series (e.g., Sousa et al., 2011). The conventional method for APS calculation relies on the assumption that the atmospheric delay is not temporally correlated. With recent SAR missions providing shorter satellite re-visit times, this assumption may no longer be appropriate. Previous studies have proposed improvements to the APS correction, including applying a global weather model (e.g., Jung et al., 2013), and accounting for the temporal variance of a pixel (e.g., Liu et al., 2011;Refice et al., 2011).
The only previous InSAR derived velocity map covering the whole Afar region was developed by Pagli et al. (2014), who used Π E -RATE to produce a displacement time series between 2005 and 2010. After removing large deformation steps associated with the Dabbahu-Manda-Hararo dyke intrusion events, Pagli et al. (2014) smoothed the time series by removing the APS, employing consistent Gaussian temporal and Butterworth spatial filters. Pagli et al. (2014) extracted 3D (east, north, vertical) velocities from ascending and descending LOS and GNSS observations on a 10-20 km resolution mesh following the method of . Surface velocities between 2005 and 2010 from Pagli et al. (2014) showed a long-term plate spreading rate of 15-20 mm/yr in Eastern Afar (relative to a stable Nubian plate), and large extension rates of  E 100 mm/yr at Dabbahu associated with the background response during the 2005-2010 Dabbahu-Manda-Hararo rifting episode. Other InSAR velocity maps within Afar have focused on individual rift segments, such as the Dabbahu-Manda-Hararo segment (Hamling et al., 2014) from 2006 to 2010, the Asal-Ghoubbet segment from 1997 to 2005 (Doubre & Peltzer, 2007(Doubre et al., 2017, and around the Tendaho Graben from 2004 to 2010 (Temtime et al., 2018).
In this study, we make use of the high temporal and spatial resolution data from the Sentinel-1 satellite to extract surface velocities from 2014 to 2019 across the whole of the Afar region. We develop and apply a small baseline methodology with spatial and temporal variance weighted filtering to improve the removal of the APS, reference the data to a stable Nubia GNSS reference frame, and calculate 3D (vertical, rift-perpendicular horizontal, and rift-parallel horizontal) velocities.

Sentinel-1 Data
We use Sentinel-1A/B acquisitions from ascending tracks 14 (014A), 87 (087A), and 116 (116A), and descending tracks 6 (006D) and 79 (079D) between October 2014 and August 2019. For processing efficiency, we divide each track into 12 (  E 250 × 250 km) frames (Table S1 in Supporting Information S1). We produce a network of geocoded unwrapped interferograms for each frame ( Figure S1 in Supporting Information S1) from single-look complex (SLC) images of each date using the LiCSAR software (González et al., 2016;Lazeckỳ et al., 2020), which automates the mass production of interferograms using GAMMA (Werner et al., 2000). To reduce noise and data size, we multi-look the SLCs at 20:4 range to azimuth looks, equating to  E 100 × 100 m pixel size. We apply a topographic correction using an SRTM (Shuttle Radar Topography Mission) 3-arc-second (  E 90 m resolution) DEM (Digital Elevation Model) (Farr & Kobrick, 2000), filter the interferograms using using a power spectrum filter (Goldstein et al., 1998), and unwrap using SNAPHU (Chen & Zebker, 2002). We manually quality check the interferogram network for each frame to remove interferograms with decorrelation, co-registration, or obvious unwrapping errors. We ensure that each epoch is connected to the network by a minimum of three interferograms by creating new interferograms as needed. Each interferogram is referenced to the mean value, excluding areas of deformation around volcanic centers.
To mitigate for atmospheric phase delay for each interferogram we compare the effectiveness of corrections from a linear trend of phase with elevation over the whole frame (e.g., Elliott et al., 2008), and the GACOS atmospheric model (e.g., Yu et al., 2017Yu et al., , 2018. For a linear phase-elevation trend correction, the mean rootmean-square (RMS) misfit for all 12 frames is reduced by 2.9 mm in comparison to the mean RMS misfit for all frames with no atmospheric correction. The GACOS correction gives a reduction in mean RMS misfit of 4.2 mm, but increases the RMS misfit in 29% of interferograms. To reduce this we follow an approach similar to Shen et al. (2019), scaling the GACOS correction for each interferogram in order to minimize the resulting RMS misfit. This improves the atmospheric correction further, producing a reduction in mean RMS misfit of 5.7 mm (see Figure S2 in Supporting Information S1). In order to account for any residual topographic atmospheric signal in each frame, we remove a linear trend of phase with elevation from each epoch, after time series filtering.

Time Series
We apply an SBAS style methodology to invert for the displacement time series at each pixel in the frame, using all interferograms where a pixel is coherent. We assess the spatial and temporal variance of the time series by first, filtering using a Laplacian filter with a temporal width of three epochs and scale factor of 3, then calculating the RMS misfit from this trend at each pixel for every epoch. We calculate the spatial distribution of RMS misfit from the time series misfits at each point, and the temporal distribution of RMS misfit from the misfits of all pixels at each epoch (see Figure S3 in Supporting Information S1). To resolve the RMS misfit value of each pixel at every epoch, we scale the spatial RMS misfit map to the temporal RMS misfit value at each epoch (see Supporting Information S1). We use these error estimates to provide weights during time series filtering, and in the inversion of filtered displacement time series for average velocities.
In order to reduce the remaining APS in the displacement time series, we filter the time series using a high-pass temporal and a low-pass spatial filter to produce the APS, which we then remove from the time series. To calculate a low-pass temporal filter, we apply a weighted linear trend with a fixed temporal width of 1 year centered on each point. To calculate the weighting for the local trend, we use the RMS misfit as a proxy for standard error, and convert the RMS misfit values into weights using the Bi-Square function where no weight is given to RMS values that exceed 6 standard deviations of the local misfits (e.g., Cleveland & Devlin, 1988). We also scale these weights by their temporal distance from the target epoch of the local time series (see Supporting Information S1). Having calculated the low-pass temporal filter, we remove it from the time series to create a high-pass temporal filter. We then apply a Gaussian spatial filter with a half-width of  E 2 km in order to resolve the APS for each epoch.
After we remove the APS, we remove a planar ramp in space and a linear trend with height to correct for any remnant long-wavelength and elevation-correlated atmospheric delay. We later restore any long-wavelength deformation removed here, using GNSS observations, which we assume correctly capture deformation on length scales of the  E 250 × 250 km Sentinel-1 frames. For each frame, we compute the average velocity at each pixel by inverting for a single linear trend through time, allowing for a constant offset. We produce a variance-covariance matrix (VCM) for each pixel, treating the temporal variation of the scaled RMS misfit as independent errors. By including the VCM in the inversion, we can quantify the uncertainty of the resulting velocities.
De Zan et al. (2015) demonstrate how a potential systematic phase-bias in interferograms with decreasing temporal baseline can influence the resulting time series. We test the magnitude of this bias by selecting consecutive 12, 24, and 36-day interferograms from frame 079D_07694_131313 covering  E 1 year (see Figure S4 in Supporting Information S1). We use a "daisy-chain" stack approach to resolve the cumulative displacements from the 12, 24, and 36 days unwrapped interferograms between December 2017 and February 2019. Any differences between these stacks indicates the presence of phase-bias and/or unwrapping errors. We find residual differences between the 12 and 24-day, and 12 and 36-day stacks of up to 50 mm, and residuals of up to 10 mm between the 24 and 36-day stacks. While this indicates that the 12-day interferograms are susceptible to a phase-bias, we find that removing the 12-day interferograms from the network effects our displacement time series by <5 mm per epoch, and our average velocities by <1 mm/yr. While we are not able to account for any bias in the 36-day interferograms, Ansari et al. (2020) indicate that the velocity bias is small in comparison to 12-day interferograms.

3D Velocities
We tie frames together within their respective tracks by sub-sampling the InSAR data points to a 5 × 5 km spacing in the overlap between frames, and 10 × 10 km spacing elsewhere, then solving for and removing a planar ramp for each frame that minimizes residuals in the along-track frame overlap regions. Removing these ramps does not bias the results as long-wavelength signals are later corrected using GNSS data. In the frame overlap region, we use the mean value of LOS velocity for each point. We find that using linear ramps to combine frames within tracks produces the fewest boundary artifacts when compared to using a single offset value calculated from the median value in the frame overlap region, or solving for a 2D quadratic function for each frame. Boundary artifacts within tracks can occur due to differences between frames in time series length, the variation in acquisition dates used, relative weighting during time series filtering, and orbital ramp removal. Although in principle it would be possible to only process and use interferograms that cover the whole along-track extent of the study region, this would require excluding several epochs where data were not acquired over the whole area, resulting in truncated time series.
To reference the LOS velocity in each track to a stable Nubian plate, we use a network of 105 GNSS stations in the Afar region to characterize long-wavelength plate motions. These data are a subset of the Geo-PRISMS community velocity field for East Africa in a Nubia-fixed International Terrestrial Reference Frame (ITRF2014), and include continuous and campaign observations acquired between 1994 and 2018 with time spans of E  2.4 years (King et al., 2019). For details of the data sources included, see King et al. (2019). We assume that these velocities are steady-state and do not vary significantly through time. We remove 32 stations in central Afar where the velocities are not steady-state, but are dominated by the ground motions associated with the 2005-2010 Dabbahu-Manda-Hararo rifting episode. As the resulting network is sparse ( Figure S5 in Supporting Information S1), with the majority of stations concentrated in Eastern Afar and few points on the Nubian and Somalian plates; we add 17 additional fabricated GNSS stations on the stable Nubian plate, with an assumed zero velocity (with uncertainties of  E 1 mm/yr and  E 2 mm/yr in the horizontal and vertical components), to help constrain the velocity field where data are sparse. We project East and North GNSS horizontal velocities into the rift-perpendicular (e.g., Hamling et al., 2014), and rift-parallel directions, oriented at 61° and −29°N respectively. From this network, we interpolate a smooth GNSS velocity field in the rift-perpendicular and rift-parallel directions over the whole Afar region at 100 × 100 m grid spacing ( Figure S5 in Supporting Information S1) using the natural neighbor algorithm (e.g., Boissonnat & Cazals, 2002). As the additional pseudo-observations define where this interpolated field reaches zero velocity, we selected these points such that they are on the Nubian plate, away from the rift border faults. We are not concerned with the precise locations, as where the interpolated velocity field reaches zero does not significantly influence the data within the Afar Rift. We estimate the error in the interpolated velocities by systematically removing each GNSS station from the network, interpolating new velocity fields in the rift-perpendicular and rift-parallel directions from the reduced network, then calculating the residual between the interpolated fields and the GNSS observation. We take the standard deviation of these residuals as the error in the rift-perpendicular and rift-parallel GNSS velocity field.
We sub-sample the InSAR LOS track velocities as previously, then extract points where there are ascending, descending, and interpolated GNSS data. We also mask points around the active rift segments so that volcanic ground deformation does not interfere with the referencing to the long-term plate motions. Using these points, we solve for the 3D velocity (rift-perpendicular, rift-parallel, vertical) at each point and a residual 2D (East, North) quadratic function for each track. We remove the respective quadratic from each InSAR track to resolve LOS velocity in a stable Nubia reference frame. The resulting LOS velocities and standard deviations are shown in Figure 2 and Figure S5 in Supporting Information S1 respectively.
In order to resolve a full 3D velocity field (vertical, rift-perpendicular horizontal, rift-parallel horizontal) at 100 × 100 m resolution, we use the smooth rift-parallel GNSS field to provide a constraint on the rift-parallel velocity at each point, as the long wavelength rift-parallel velocities are small in comparison to the rift-perpendicular and vertical velocities. We include this constraint with the ascending and descending LOS InSAR observations to calculate 3D velocities at each point using a least squares inversion (Hussain et al., 2016;Weiss et al., 2020;Wright et al., 2004 ). We weight the inversion and resolve uncertainties by including a diagonal VCM using the previously calculated variance at each point. (Figure 3), discussed in Section 5, show both the long-term plate motion and surface deformation associated with magmatism and transient tectonics. As we use the interpolated GNSS velocity field as an additional constraint in the rift-parallel direction, the resulting velocity field is not sensitive to short-wavelength rift-parallel deformation, which could be locally significant (Smittarello et al., 2016). As the rift-parallel error estimates also derive from the interpolated GNSS data, they are lower than the rift-perpendicular errors, which we calculate using only the InSAR observations to retain high spatial resolution (see Figure 3).

Plate Motions and Uncertainties
Our 2014-2019 horizontal velocity maps (Figure 3) show the rift-perpendicular extension over the Afar rift zone at rates of up to 25  E 5 mm/yr, with negligible long-wavelength motions in the rift-parallel direction. We also observe the gradual North to South transfer of extension from the Red Sea Rift into the Afar rift zone between  E 13-16°N, indicative of the rotation of the Danakil micro-plate relative to the stable Nubian plate (e.g., Kidane, 2016;Manighetti et al., 2001;Viltres et al., 2020).
Areas of noise up to  E 10 mm/yr over the Ethiopian Highlands region on the Nubain plate, are highlighted in the standard deviation maps shown in Figure 3. The regions of high error in T087A, and the northern-most portion of T014A (see Figure S6 in Supporting Information S1), are a result of the shorter time series length in these regions producing more uncertainty in the long-term velocity estimates. Elsewhere, errors of up to  E 5 mm/yr are likely due to uncorrected atmospheric delays, and artifacts over track boundaries, where we are unable to account for small LOS velocity variations between overlapping tracks. These discontinuities 10.1029/2020JB021542 8 of 18 may arise from variations in atmospheric filtering between overlapping frames, and in GNSS referencing between overlapping tracks. The relatively uniform errors in the rift-parallel direction are due to the GNSS velocity field being used to constrain the velocities in this component, where the velocities are small in comparison to the vertical and rift-perpendicular components.
Using the overlap region between each of the tracks, we test the internal consistency of our LOS InSAR data using cross-validation by excluding each track in turn from the 3D inversion, projecting the resulting velocity field into the LOS of the excluded track, and calculating the residuals. The RMSE of the residuals in each track overlap region ( Figure S7 in Supporting Information S1) range from 2.7 to 4.2 mm/yr. These are consistent with the combination of the uncertainties in our LOS velocities (1.5-6.6 mm/yr) and the propagated uncertainties of the predicted LOS velocities from our 3D velocity field with one track excluded (2.8-2.9 mm/yr), and support our estimate of uncertainties shown in Figure 3.
We also isolate the contribution of the LOS InSAR velocities to the rift-perpendicular and rift-parallel horizontal velocity maps by subtracting the interpolated GNSS velocities from the final 3D velocity map. We only test the horizontal directions of the velocity field here as we do not use GNSS data to constrain the vertical direction. As shown in Figure S8 in Supporting Information S1, velocities in the rift-perpendicular  1 tracks T087A, T014A, T116A,  T079D, and T006D. LOS velocities are referenced to a stable Nubia reference frame using long-term plate motions from the regional Global Navigation Satellite System network (King et al., 2019). Arrows indicate the track look directions. Volcano-tectonic segments and key faults are shown as black dashed outlines and black solid lines respectively. 9 of 18 direction use contributions from both InSAR and GNSS, while the rift-parallel velocities are constrained only by the GNSS component as described previously. In the rift-perpendicular direction, the InSAR component is dominated by local tectono-magmatic signals and residual atmospheric and processing artifact signals, indicating that our referencing of LOS InSAR velocities to far-field plate motions is robust.

Rift Extension and Focusing
Profiles taken across the rift zone highlight the focusing of extension in Afar. Profiles covering the Alid graben, at northern-most tip of the Afar rift zone (Figures 4b and 5b), show that a broad uplift and extensional signal of up to  E 20 mm/yr is centered within  E 10-15 km of the rift axis. This may be indicative of deep magmatic intrusion in an area with the smallest background extension rates throughout Afar, but comparable to extension rates at the active volcanic islands at the southern end of the oceanic Red Sea Rift (Eyles et al., 2018). Uncertainty estimates in this region are also significantly higher than elsewhere in Afar and as such we do not investigate this signal any further. Profiles traversing the Erta 'Ale segment in the Danakil Depression (Figures 4c and 5c) highlight that the majority of extension here is focused into a region within  E 15-20 km of the rift axis. Outside of this region, the rate of extension does not significantly vary, with velocities in agreement with long-term GNSS observations (Figure 4c). We also observe subsidence on the Erta 'Ale segment between the Erta 'Ale and Alu-Dalafilla volcanoes (Figure 4c), which could be linked to magma withdrawal associated with the 2017-2019 eruption at Erta 'Ale (Moore et al., 2019;Xu et al., 2017Xu et al., , 2020. Profiles between the Erta 'Ale volcano-tectonic segment and the Dabbahu-Manda-Hararo volcano-tectonic segment (Figure 5c) show that extension in this region is shared between the Alayta volcano-tectonic segment and the Tat 'Ale volcano-tectonic segment, and focused to within  E 10-20 km of the rift segments. Profiles covering the Dabbahu-Manda-Hararo volcano-tectonic segment (Figures 4d and 5d) also show that the long-term extension is concentrated near the rift axis, with only small variations in rift-perpendicular  Figure 6. The location of modeled deformation sources for a 0.9-1.3 km deep sill (Okada, 1985) at Dallol (F, Figure S10 in Supporting Information S1) and a 5.5-6.8 km deep point source (Mogi, 1958) at Nabro (G, Figure S11 in Supporting Information S1) are shown as black outlines. Profiles over southern central Afar and the Asal-Ghoubbet segment (Figures 4e and 5e) show a more distributed pattern of extension with an initial increase in rift-perpendicular velocities 70-140 km to the SW of the rift axis near the Tendaho Graben, before velocities stabilize at  E 20 mm/yr on the Danakil micro-plate within 10-20 km to the NE of the rift axis. The distribution of strain shows good agreement with GNSS observations in Figure 4e, with the differences likely occurring due to the distances between the profiles and GNSS sites (see Figure 4a) particularly for the Asal-Ghoubbet segment, where our rift-perpendicular Figure 5. Map of 21 10 km wide rift-perpendicular velocity (relative to stable Nubia, positive toward 61°N) profiles over the Afar rift zone (a), with simplified rift segments shown by white dashed outlines, and simplified faults by solid white lines. The velocity profiles cover northern Afar (b), the Erta 'Ale and Tat 'Ale segments (c), the Dabbahu-Manda-Hararo segment (d), and the Manda-Inakir and Asal-Ghoubbet segments (e). The region on each profile where the majority of extension is accommodated, is indicated by the solid portion of the profile lines in (a) and (b-e), with the rest of the profile marked as dotted lines. This region of extension is approximated for each subset by the dark gray shaded portions of (b-e). The standard deviation of rift-perpendicular velocities varies from  E 2-7 mm/yr (see Figure 3). The mean elevation profile for each group of profiles is shown as a solid black line. Profile distances are measured arbitrarily, relative to where the profile coincides with a magmatic segment, with positive toward the NE. Zero profile distance is marked by black circles on (a), and dashed black lines on (b-e). These profiles are shown individually with error estimates in Figure S9 in Supporting Information S1. velocities vary strongly along the rift. In southern Afar, extension between 2014 and 2019 may be largely accommodated by tectonic rather than magmatic mechanisms, with strain being distributed across a sequence of horst and graben structures (e.g., Manighetti et al., 2001;Tapponnier et al., 1990). Previous studies have also shown that strain in southern central Afar is distributed across a broad region of the rift, with InSAR derived velocities from 2003-2010 (Doubre et al., 2017(Pagli et al., 2014 indicating that along-profile extension is distributed over  E 50-100 km and  E 70-120 km respectively. The detailed GNSS observations of Doubre et al. (2017) demonstrate how this extension is accommodated within the central Afar fault structures, where extension is accommodated by tectonic processes (Manighetti et al., 1998(Manighetti et al., , 2001. Pagli et al. (2019) suggest that the elevated strain and seismicity in central Afar demonstrates strain transfer between the Dabbahu-Manda-Hararo segment and Asal-Ghoubbet segment. Doubre and Peltzer (2007) and Doubre et al. (2017) also show small steps in extension close to the Asal-Ghoubbet segment with an increase of  E 2-6 mm/yr over the rift segment, in agreement to the step shown in Figure 4e of  E 1-7 mm/yr.
Our results indicate that at segments with current active magmatism in central and northern Afar (Erta 'Ale and Dabbahu-Manda-Hararo rift segments), extension is largely focused to within  E 15-30 km of the rift axis; while in southern Afar, between the Tendaho Graben and the Manda-Inakir and Asal-Ghoubbet rift segments, extension may be distributed over 80-160 km. The broad distribution of strain in southern Afar is comparable with previous GNSS and InSAR based studies including Kogan et al. (2012), Pagli et al. (2014), and Doubre et al. (2017). Kogan et al. (2012) show that extension in southern Afar occurs over  E 175 km, but also suggest that extension becomes more distributed with rift development, although their results are based on one profile through the amagmatic portion of the Afar rift zone and may also be influenced by the proximity to the triple junction. In contrast, our results suggest a possible increase in strain focusing with rift maturity during late-stage continental breakup, in keeping with strain localization assisting the transition into oceanic spreading centers (e.g., Manighetti et al., 1998), and with recent findings in the MER which suggest a northwards increase in the extension rate (Temtime, 2021). Our findings are also in keeping with other studies of rift development in Afar (e.g., Bastow et al., 2018;Keir et al., 2013), which also suggest an increase in localization in northern Afar, with significant plate thinning and extrusive magmatism. Figure 4 highlights the surface deformation at Dallol (4F) and Nabro (4G) volcanoes, and at the Dabbahu-Manda-Hararo segment (4I), where localized deformation, likely associated with magma migration, is visible. As magmatic deformation may not be steady in time, we look at time series for points located in the middle of these centers. Time series of vertical displacements at the Dallol, Nabro, Dabbahu, and Manda-Hararo volcanic centers show that the deformation is reasonably linear from 2014 to 2019, indicating that the velocities displayed in Figures 4f-4i are approximately representative of 2014-2019 ground motions. For Erta 'Ale volcano, we select a point  E 2 km to the north of the summit caldera in order to avoid the step surface deformation associated with a dyke intrusion in January 2017 (Moore et al., 2019;Xu et al., 2017Xu et al., , 2020. Following this intrusion the Erta 'Ale edifice shows linear subsidence at a rate of 15  E 4 mm/yr ( Figure 6). Our results suggest a continuation of the inter-dyking period at the Asal-Ghoubbet segment (e.g., Cattin et al., 2005;, with negligible vertical deformation, although we are unable to resolve any local rift-parallel deformation.

Magmatic Deformation
At Dallol volcano, at the northern end on the Erta 'Ale segment, we observe a high rate of subsidence of up to 45  E 4 mm/yr, with negligible horizontal movement. The subsidence signal is focused on the central cone at Dallol. We model this signal using the Markov-Chain Monte-Carlo Geodetic Bayesian Inversion Software (GBIS) (Bagnardi & Hooper, 2018). For the T014A and T079D LOS deformation between 2014 and 2019, we test source geometries including a point pressure source (Mogi, 1958), a planar dislocation (Okada, 1985), and a penny-shaped crack (Fialko et al., 2001). We find that a  E 1 × 2 km horizontal sill at 0.9-1.3 km depth with  E 0.27 m of contraction gives the lowest residual RMS misfit ( Figure S10 in Supporting Information S1). At Nabro volcano, we observe edifice uplift and extension at rates of up to 12  E 3 mm/yr, combined with the subsidence of lava flows from the 2011 eruption (J. E. Hamlyn et al., 2014), and a highly localized subsidence and contraction signal of up to 14  E 3 mm/yr at the center of the Nabro caldera. This uplift of the Nabro edifice shows a change from InSAR observed subsidence of 150-200 mm/yr from 2011 to 2012, after the 2011 eruption (J. Hamlyn et al., 2018). As our vertical displacement time series ( Figure 6) indicates that the uplift at Nabro is linear between 2014 and 2019, we suggest that this post-eruption edifice subsidence must have stopped between 2012 and 2014. We model T006D, T014A, and T079D LOS observations at Nabro volcano between 2014 and 2019 using a point pressure source (Mogi, 1958), and penny-shaped crack (Fialko et al., 2001) using GBIS (Bagnardi & Hooper, 2018). Figure S11 in Supporting Information S1 shows the optimal Mogi source at 5.5-6.8 km depth with a volume increase of 7-11 × 6 10 E  work is required to model the long-term response to the 2005-2010 rifting episode at the Dabbahu-Manda-Hararo segment, incorporating a viscous response and continued magma movement by combining the InSAR time series of Pagli et al. (2014), with the data presented here.

Conclusions
We develop Sentinel-1 displacement time series at 100 × 100 m resolution between 2014 and 2019 over three ascending and two descending tracks, covering the whole Afar rift zone. We implement a RMS misfit weighted APS correction to clean the time series, and produce average velocity maps for each frame. Using GNSS observations of long-term plate motions (King et al., 2019), we reference the InSAR velocities to the stable Nubian plate, and convert LOS into 3D velocities (vertical, rift-perpendicular, rift-parallel).
We are able to resolve deformation at individual volcanic centers, with subsidence of 45  E 4 mm/yr at Dallol volcano, consistent with the deflation of a shallow sill at 0.9-1.3 km depth. We also show that edifice uplift at Nabro volcano of 12  E 3 mm/yr is sourced from a magma chamber at 5.5-6.8 km depth, consistent with the source of post-eruption subsidence observed between 2011 and 2012 (J. Hamlyn et al., 2018). Pagli et al. (2014) and Hamling et al. (2014) identify vertical and rift-perpendicular horizontal surface velocities between 2006 and 2010 of 80-240 mm/yr and 110-180 mm/yr, respectively associated with a background post-rift response to the initial 2005 dyking episode at the Dabbahu-Manda-Hararo segment. We show that this response is ongoing between 2014 and 2019, but at lower rates of 33  E 4 mm/yr and 37  E 4 mm/yr for uplift and rift-perpendicular extension respectively. We suggest that this  E 15 years response to the 2005 dyke intrusion is indicative of continued magma movement and/or time-dependant viscous processes within the crust below the rift segment.
We resolve the long-wavelength extension over the Afar rift zone with rift-perpendicular velocities of up to 25  E 5 mm/yr, with negligible motions in the rift-parallel direction. From cross-rift profiles, we find that extension is largely focused to within  E 15-30 km of the rift axis on the active magmatic rift segments in northern Afar, while strain in central and southern Afar is distributed across 80-160 km of the rift. This trend of increased focusing of extension into northern Afar is consistent with strain localization assisting the transition into oceanic spreading centers.

Data Availability Statement
All Sentinel-1 data is sourced from the European Union Copernicus Programme. We perform data processing on the JASMIN facility, operated by the Centre for Environmental Data Analysis (CEDA), and post-processing and figure generation using MATLAB and GMT. Processed LiCSAR interferograms are available in the Centre for Environmental Data Analysis (CEDA) archive here: http://data.ceda.ac.uk/ neodc/comet/data/licsar_products. Final rift-perpendicular, rift-parallel, and vertical surface velocity and estimated variance maps are also available in the CEDA archive here: http://dx.doi.org/10.5285/ ac43cee2bf5e4942970492209ba95e49