Multi-Baseline InSAR: Theory & Practice

• Multi-Baseline InSAR is a remote sensing technique that integrates multiple SAR images acquired over varying baselines to accurately retrieve terrain deformation, digital elevation models, and elevation spectra.
• It exploits baseline diversity to overcome the ill-posed height retrieval problem inherent in single-baseline methods, enabling techniques such as SBAS, UAV fusion, and tomographic inversion.
• Recent advancements include recursive phase linking, compressive sensing-based tomographic inversion, and optimized UAV communication strategies that improve scalability and accuracy.

Searching arXiv for recent and foundational papers on multi-baseline InSAR to ground the article in current literature. Search query: "multi-baseline InSAR tomography phase linking SBAS UAV InSAR arXiv" Multi-baseline interferometric synthetic aperture radar (MB-InSAR) denotes the family of SAR interferometric methods that aggregate multiple acquisitions from different times, viewing geometries, or spatially separated antennas in order to estimate deformation, digital elevation models (DEMs), or elevation-resolved reflectivity. Relative to single-baseline InSAR, MB-InSAR exploits baseline diversity to overcome the ill-posed problem of single-baseline height retrieval and to support deformation time-series inversion, phase linking, and SAR tomography. In the contemporary literature, the term covers repeat-pass stacks of co-registered single-look complex (SLC) images, bistatic and multi-master acquisitions such as TanDEM-X, single-pass tandem systems, automotive synthetic apertures, and UAV swarms with communication-assisted processing (Hu et al., 2023, Lahmeri et al., 15 Jul 2025).

1. Measurement models and inverse formulation

A standard MB-InSAR formulation starts from a stack of $N$ co-registered SLC SAR images $y_1,\dots,y_N$ acquired at increasing times over the same scene. The basic observables are the interferometric phases

$\phi_{i,j}=\angle(\overline y_i\,y_j),$

or, equivalently, a consistent time series of complex phase factors $\{\theta_k=e^{j\psi_k}\}$ such that $\phi_{i,j}\approx \psi_j-\psi_i$. In weighted least squares, one seeks $\{\psi_k\}$ by minimizing

$$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$

with $w_{i,j}$ chosen proportional to coherence or derived from a full temporal-covariance model (Zan, 2020).

A covariance-based formulation makes the same structure explicit. For a stack of $l$ SAR images, the complex returns in a homogeneous neighborhood are modeled as $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$, with the physically consistent covariance

$y_1,\dots,y_N$0

where $y_1,\dots,y_N$1 and $y_1,\dots,y_N$2 lies on the complex torus $y_1,\dots,y_N$3. Phase-linking then becomes projection of a plug-in covariance estimate onto that model, typically by minimizing a Frobenius-norm or Kullback–Leibler distance (Hajjar et al., 13 Feb 2025).

When the goal is topography rather than phase history, the phase-to-height relation enters directly. In dual-baseline or multi-baseline UAV-InSAR, the interferometric phase difference on baseline $y_1,\dots,y_N$4 is written as

$y_1,\dots,y_N$5

leading to the height of ambiguity

$y_1,\dots,y_N$6

and a baseline-specific height error standard deviation determined by coherence and the number of looks (Lahmeri et al., 2024). In generalized multi-UAV fusion, each interferogram receives weight $y_1,\dots,y_N$7, and the final DEM estimate is the weighted mean of the baseline-specific height estimates (Lahmeri et al., 15 Jul 2025).

These formulations define the central MB-InSAR inverse problems: recovering a phase history consistent across many interferograms, recovering a deformation time series on acquisition dates, or recovering a height/elevation profile from measurements acquired over multiple baselines.

2. Multi-temporal networks and SBAS deformation retrieval

In deformation monitoring, MB-InSAR commonly appears as a network inversion problem over acquisition dates. The Small Baseline Subset (SBAS) method is a special case of MB-InSAR in which only interferograms with limited temporal and perpendicular baselines are retained in order to reduce decorrelation. The residual interferometric phase can be modeled as the sum of deformation, residual topography, atmospheric contribution, DEM-related residuals, and noise. After topographic correction and atmospheric mitigation, the stack is written in linear form as

$y_1,\dots,y_N$8

where $y_1,\dots,y_N$9 is the vector of unwrapped interferometric phases, $\phi_{i,j}=\angle(\overline y_i\,y_j),$0 is the displacement time series, and each row of $\phi_{i,j}=\angle(\overline y_i\,y_j),$1 contains $\phi_{i,j}=\angle(\overline y_i\,y_j),$2 at the master date and $\phi_{i,j}=\angle(\overline y_i\,y_j),$3 at the slave date (Fatholahi et al., 2021).

The admissible interferogram set is selected by thresholds on temporal and perpendicular baselines. In the Marun oilfield study, typical values were $\phi_{i,j}=\angle(\overline y_i\,y_j),$4 days and $\phi_{i,j}=\angle(\overline y_i\,y_j),$5 m. The selected interferograms form a graph whose nodes are acquisition dates and whose edges are the retained interferograms; connectivity is required so that every date is reachable through a chain of small-baseline pairs. In that case, 10 Envisat ASAR images yielded 45 possible pairs, of which 22 satisfied both thresholds, giving a connected network on 10 nodes (Fatholahi et al., 2021).

Inversion is usually posed as regularized weighted least squares,

$\phi_{i,j}=\angle(\overline y_i\,y_j),$6

with coherence-derived weights $\phi_{i,j}=\angle(\overline y_i\,y_j),$7 and a temporal differencing operator $\phi_{i,j}=\angle(\overline y_i\,y_j),$8 enforcing smoothness. Independent unwrapping of each interferogram can be stabilized by exploiting graph redundancy: around any closed loop in the interferogram network, the phase sum should be zero, so cycle slips can be detected and corrected through multi-baseline consistency checks (Fatholahi et al., 2021).

The Marun application illustrates the operational role of this framework. Using 10 Envisat ASAR images and 22 interferograms, the mean velocity map in the line-of-sight direction revealed a maximum subsidence on the order of $\phi_{i,j}=\angle(\overline y_i\,y_j),$9 mm per year, and comparison with oil production rate showed good agreement. The result is characteristic of MB-InSAR time-series analysis: the method does not treat interferograms independently, but uses a constrained network to separate deformation from nuisance terms and to estimate a temporally consistent displacement history (Fatholahi et al., 2021).

3. Covariance-driven phase linking and recursive processing

A second major MB-InSAR branch focuses on phase linking across long stacks. In a full-covariance benchmark, the temporal covariance matrix $\{\theta_k=e^{j\psi_k}\}$0 has elements $\{\theta_k=e^{j\psi_k}\}$1, and the maximum-likelihood phase estimate solves

$\{\theta_k=e^{j\psi_k}\}$2

This formulation underlies full-covariance algorithms such as EMI, but it requires access to all past data and inversion or eigendecomposition of an $\{\theta_k=e^{j\psi_k}\}$3 matrix (Zan, 2020).

To reduce this burden, progressive phase estimation replaces the full stack by a recursive pair of complex buffers: a running reference $\{\theta_k=e^{j\psi_k}\}$4 and a stable reference $\{\theta_k=e^{j\psi_k}\}$5. At each incoming acquisition $\{\theta_k=e^{j\psi_k}\}$6, the algorithm forms the short-term interferogram $\{\theta_k=e^{j\psi_k}\}$7, updates the running reference through an AR(1) recursion,

$\{\theta_k=e^{j\psi_k}\}$8

calibrates it against the stable reference using

$\{\theta_k=e^{j\psi_k}\}$9

and then accumulates $\phi_{i,j}\approx \psi_j-\psi_i$0. The parameter $\phi_{i,j}\approx \psi_j-\psi_i$1 controls memory versus responsiveness: $\phi_{i,j}\approx \psi_j-\psi_i$2 retains long-term energy but risks drift, whereas $\phi_{i,j}\approx \psi_j-\psi_i$3 favors short-term coherence but risks noise in long-term phase (Zan, 2020).

The operational implication is substantial. Only two complex rasters of scene size are needed in memory, there is no need to retrieve or store the full image stack or a full covariance matrix, and per-acquisition complexity is $\phi_{i,j}\approx \psi_j-\psi_i$4. The scheme yields new $\phi_{i,j}\approx \psi_j-\psi_i$5 as soon as each SLC arrives, without batch reprocessing, and can use per-interferogram coherence as an online quality metric (Zan, 2020).

Performance analyses in simulation and on real data show why this recursive formulation is significant. For a C-band Sentinel-1 decorrelation model over $\phi_{i,j}\approx \psi_j-\psi_i$6 up to $\phi_{i,j}\approx \psi_j-\psi_i$7 days, uncalibrated recursion exhibited approximately $\phi_{i,j}\approx \psi_j-\psi_i$8 mm/yr drift, whereas drift-calibrated recursion saturated at $\phi_{i,j}\approx \psi_j-\psi_i$9 mm and had a standard deviation almost identical to EMI. In Sicily data from 2014–2018, the uncalibrated version showed approximately $\{\psi_k\}$0–$\{\psi_k\}$1 mm/yr drift versus EMI, while the calibrated version had residual bias of approximately $\{\psi_k\}$2–$\{\psi_k\}$3 mm and a quasi-periodic annual oscillation attributed to vegetation effects; interferogram quality remained comparable to EMI, with slightly higher short-term coherence due to adaptive re-phasing (Zan, 2020).

Sequential covariance fitting addresses the same scalability issue from another direction. Seq-COFI-PL treats phase linking as optimization on the complex torus and solves the resulting quadratic objective by Majorization–Minimization. If $\{\psi_k\}$4 images are partitioned into $\{\psi_k\}$5 old acquisitions and a newly acquired block of $\{\psi_k\}$6 images, the full covariance is block-partitioned and only the new phases are estimated while the old block is kept fixed. This changes the dominant cost from batch $\{\psi_k\}$7 in the KL case to $\{\psi_k\}$8, with analogous savings for the Frobenius-norm formulation (Hajjar et al., 13 Feb 2025).

Simulation with $\{\psi_k\}$9, $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$0, $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$1 showed identical mean-squared error between Seq-COFI-PL and batch COFI-PL for the $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$2–$$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$3 interferometric phase difference over a range of spatial samples, with no accumulation of bias or noise increase across successive block integrations. On real Sentinel-1 data over Mexico City, Seq-COFI-PL reproduced the longest-baseline interferogram at the same quality as batch COFI-PL while running approximately $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$4 faster for KL and approximately $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$5 faster for Frobenius; a recent sequential MLE-PL was roughly four times slower, and a mini-stack PCA approach was noisier (Hajjar et al., 13 Feb 2025).

4. Tomographic MB-InSAR and 3-D urban reconstruction

In SAR tomography, MB-InSAR is used to recover the elevation distribution of scatterers inside a range–azimuth pixel. The continuous forward model expresses each interferometric measurement as a Fourier sample of the elevation reflectivity profile: $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$6 After discretization into $$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$7 elevation bins, the inverse problem becomes

$$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$8

or, in the multi-master notation,

$$J(\psi)=\sum_{i<j} w_{i,j}\bigl|\phi_{i,j}-(\psi_j-\psi_i)\bigr|^2,$$9

where $w_{i,j}$0 or $w_{i,j}$1 is the discretized elevation-dependent reflectivity, and the sensing matrix entries are complex exponentials of the effective baseline-wavenumber differences (Shi et al., 2020, Shi et al., 2023).

Two inversion families dominate. A Wiener-type SVD spectral estimator uses

$w_{i,j}$2

while a compressive-sensing estimator solves

$w_{i,j}$3

In the multi-master TanDEM-X framework, a practical hybrid is used: a low-order detector first estimates the number of scatterers per pixel, then $w_{i,j}$4-minimization is applied only where multiple scatterers are present (Shi et al., 2020, Shi et al., 2023).

Because very small stacks provide insufficient raw $w_{i,j}$5, both Munich studies integrated non-local filtering before tomographic inversion. For each central pixel, a $w_{i,j}$6 patch is compared with every $w_{i,j}$7 patch in a $w_{i,j}$8 search window, similarity weights are computed from a log-likelihood distance of joint intensity/phase statistics, and a weighted maximum-likelihood estimate of interferometric parameters is formed. The filtered complex interferogram then replaces the raw measurement vector in the tomographic inversion. In the micro-stack setting of $w_{i,j}$9–$l$0 interferograms, this boosts effective SNR by $l$1–$l$2 dB without significant loss of spatial resolution (Shi et al., 2020).

The central result of this line of work is that a conventionally data-hungry MB-InSAR task can be pushed into the micro-stack regime under favorable conditions. The Munich experiments used five TanDEM-X bistatic interferograms acquired from July 2016 to July 2017, with maximal elevation aperture $l$3 m. Preprocessing included coregistration, deramping, flattened-phase subtraction, spectral estimation, model selection by AIC/BIC/MDL, robust height fusion, and geocoding to UTM (Shi et al., 2020).

Monte Carlo analyses quantified the attainable precision. For single scatterers, both SVD and CS approached the CRLB

$l$4

asymptotically once $l$5 dB. For double scatterers representing facade-ground mixtures, SVD resolved only $l$6, whereas CS achieved super-resolution down to $l$7 at $l$8 dB and $l$9 (Shi et al., 2020).

On real data, the building-height results were city-scale rather than anecdotal. Against airborne LiDAR, the relative top-bottom height difference errors for nine representative buildings were $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$0 m for TomoSAR, with mean absolute difference approximately $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$1–$\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$2 m and $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$3 m, whereas the TanDEM-X raw DEM showed errors of $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$4–$\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$5 m. Across $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$6 buildings, $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$7 were within $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$8 m and $\tilde x^i\sim\mathcal{CN}(0,\widetilde\Sigma)$9 within $y_1,\dots,y_N$00 m of LiDAR; after truncating outliers $y_1,\dots,y_N$01 m, the remaining $y_1,\dots,y_N$02 buildings had $y_1,\dots,y_N$03 m (Shi et al., 2020, Shi et al., 2023).

A frequent misconception is that plausible urban reconstruction always requires more than twenty interferograms. The Munich studies do state that large stacks are conventionally necessary, but they also show that extremely small bistatic stacks can be made viable when non-local filtering is integrated into inversion, while preserving the known limitations: $y_1,\dots,y_N$04 scatterers per pixel remains a lower bound, detection of more than two scatterers is fragile for $y_1,\dots,y_N$05–$y_1,\dots,y_N$06, and low-coherence or vegetation-dominated areas remain difficult (Shi et al., 2020, Shi et al., 2023).

5. Single-pass architectures, automotive MB-InSAR, and UAV-based systems

Single-pass MB-InSAR pursues the same height-retrieval objective under a different coherence regime. Tandem Dual-Antenna SAR Interferometry (TDA-InSAR) uses two co-orbiting satellites, each carrying two antennas, to obtain three independent perpendicular baselines in a single pass. After flat-Earth removal, the interferometric phase on baseline $y_1,\dots,y_N$07 is

$y_1,\dots,y_N$08

and the corresponding unwrapped height estimate is

$y_1,\dots,y_N$09

The system is designed around an asymptotic 2D+1D phase-unwrapping sequence: a real short-baseline interferogram is spatially unwrapped first, then integer ambiguities are estimated for a medium-baseline interferogram and finally for a long-baseline interferogram (Hu et al., 2023).

TDA-InSAR formalizes two design metrics: expected relative height precision and successful phase-unwrapping rate. For a perfectly unwrapped long-baseline interferogram,

$y_1,\dots,y_N$10

while the overall success rate is approximated from the ambiguity-estimation probabilities of the medium- and long-baseline stages. In simulation for built-up targets with coherence approximately $y_1,\dots,y_N$11, Configuration 2 with $y_1,\dots,y_N$12 m, $y_1,\dots,y_N$13 m, and $y_1,\dots,y_N$14 m yielded $y_1,\dots,y_N$15 m and success rate approximately $y_1,\dots,y_N$16. For forest canopy at L-band, the point-scatterer RMSE was $y_1,\dots,y_N$17 m, the mean error was $y_1,\dots,y_N$18 m, and coverage was $y_1,\dots,y_N$19; a two-sample F-test showed no significant loss versus a 29-image repeat-pass 3D phase-unwrapping reference (Hu et al., 2023).

Automotive MB-InSAR moves the same principles to a low-cost mobile platform. A TI AWR1243BOOST front-end with $y_1,\dots,y_N$20 transmit and $y_1,\dots,y_N$21 receive elements forms a $y_1,\dots,y_N$22-element virtual array arranged in two elevation layers with vertical offset $y_1,\dots,y_N$23, yielding $y_1,\dots,y_N$24 independent vertical baselines. Interferometric phase is measured per pixel as

$y_1,\dots,y_N$25

the average phase across baselines is converted to elevation angle, and Cartesian coordinates are recovered from the range, azimuth, and elevation geometry (Kabuli et al., 14 Jan 2025).

The processing chain is explicitly tailored to automotive constraints: FMCW chirps with $y_1,\dots,y_N$26 MHz bandwidth, TDM-MIMO coding at $y_1,\dots,y_N$27 GHz, fast back-projection over a $y_1,\dots,y_N$28 m aperture, per-pixel unwrapping assisted by the bound $y_1,\dots,y_N$29, SNR thresholding at $y_1,\dots,y_N$30 dB, a baseline-consistency filter removing pixels whose phase varies by more than $y_1,\dots,y_N$31 across baselines, and geometric cutoffs that discard implausible returns. Controlled-chamber validation reported a $y_1,\dots,y_N$32 cm reflector measured as $y_1,\dots,y_N$33 cm, a $y_1,\dots,y_N$34 cm reflector measured as $y_1,\dots,y_N$35 cm, a $y_1,\dots,y_N$36 cm ground reflector measured as $y_1,\dots,y_N$37 cm, and overall RMSE approximately $y_1,\dots,y_N$38 cm. In on-vehicle tests, urban and agricultural scenes showed correct absolute elevations for cars, vegetation, and buildings, while elevation accuracy degraded with range but still resolved human-scale features within $y_1,\dots,y_N$39 cm (Kabuli et al., 14 Jan 2025).

UAV-based MB-InSAR incorporates an additional systems problem: radar data must be offloaded to a ground station in real time while formation geometry is chosen to optimize DEM precision. In a communication-assisted dual-baseline system with three UAV-borne SARs, two independent DEMs are fused by inverse-variance weighting. Under worst-case decorrelation, the fused height-error upper bound is

$y_1,\dots,y_N$40

Joint optimization of UAV formation and communication power by AO+SCA reduced the worst-case height estimation error by approximately $y_1,\dots,y_N$41 relative to a classical single-baseline UAV-InSAR benchmark, by a minimum of approximately $y_1,\dots,y_N$42 relative to a fixed-master dual-baseline benchmark, and by more than $y_1,\dots,y_N$43 relative to static power allocation; the static-power scheme became infeasible beyond $y_1,\dots,y_N$44 Mbps (Lahmeri et al., 2024).

A more general multi-UAV extension considers $y_1,\dots,y_N$45 SAR platforms and $y_1,\dots,y_N$46 baselines, with joint optimization of formation, along-track velocity, and FDMA communication powers using a co-evolutionary particle-swarm framework. In simulation, the method converged to feasibility within approximately $y_1,\dots,y_N$47 outer generations and reached $y_1,\dots,y_N$48 cm for $y_1,\dots,y_N$49; height error stayed below $y_1,\dots,y_N$50 cm for $y_1,\dots,y_N$51 m, gains were dramatic from $y_1,\dots,y_N$52 UAVs, and sub-decimeter accuracy was achievable already at $y_1,\dots,y_N$53–$y_1,\dots,y_N$54, with saturation beyond $y_1,\dots,y_N$55 (Lahmeri et al., 15 Jul 2025).

6. Performance trade-offs, misconceptions, and research directions

Across these formulations, MB-InSAR is governed by a recurrent trade-off between baseline diversity and coherence. Long baselines improve vertical sensitivity and reduce $y_1,\dots,y_N$56, but they also increase decorrelation and phase-unwrapping risk; short baselines preserve coherence and enlarge the height of ambiguity, but limit precision. TDA-InSAR states this as a baseline optimization problem over success rate, height of ambiguity, and expected height precision, while multi-UAV optimization expresses the same balance through explicit constraints on $y_1,\dots,y_N$57, coverage, communication rate, and energy budget (Hu et al., 2023, Lahmeri et al., 15 Jul 2025).

A second recurring theme is that “more data” does not have a single meaning. In TomoSAR micro-stacks, increasing SNR can yield more benefit than adding another interferogram when $y_1,\dots,y_N$58 is very small, and the relevant asymptotic threshold in the Munich simulations was $y_1,\dots,y_N$59 dB (Shi et al., 2020). In covariance-based phase linking, batch access to all past images improves modeling flexibility but creates $y_1,\dots,y_N$60 memory and $y_1,\dots,y_N$61 computation burdens, which progressive recursion and sequential covariance fitting explicitly try to avoid (Zan, 2020, Hajjar et al., 13 Feb 2025).

Another common misconception is that recursive or sequential MB-InSAR must sacrifice long-term accuracy. The published evidence is narrower than such a general statement, but it does show that progressive phase estimation with drift calibration remained stable over years and stayed within a few millimeters of a full-covariance benchmark, and that Seq-COFI-PL matched batch COFI-PL quality for the tested Sentinel-1 cases while reducing computation (Zan, 2020, Hajjar et al., 13 Feb 2025).

The literature also delineates method-specific failure modes. In SBAS, the interferogram network must remain connected and low-coherence pixels are down-weighted or discarded (Fatholahi et al., 2021). In micro-stack TomoSAR, detection of more than two scatterers is fragile, low-coherence roofs and narrow streets may fail, baseline clustering reduces vertical aperture, and the white-scattering stationarity assumption may break down over vegetation or non-homogeneous targets (Shi et al., 2020, Shi et al., 2023). In automotive MB-InSAR, the vertical aperture $y_1,\dots,y_N$62 limits single-baseline height sensitivity, the single-source-per-range-azimuth-bin approximation fails for closely spaced scatterers, and TDM coding restricts vehicle speed to $y_1,\dots,y_N$63 m/s (Kabuli et al., 14 Jan 2025). In progressive phase estimation, drift calibration is ineffective in the first approximately $y_1,\dots,y_N$64 days because the algorithm is causal, and a single long-term component may be insufficient when multiple evolving scatterer classes are present (Zan, 2020).

The stated research directions are correspondingly diverse. For recursive phase estimation, proposed extensions include ARMA or higher-order predictors, multiple stable references $y_1,\dots,y_N$65, adaptive $y_1,\dots,y_N$66-estimation, and integration with spatial-adaptive coherence estimation (Zan, 2020). For tomographic urban mapping, proposed directions include phase closure relations in multi-master stacks, fusion with optical stereo or LiDAR, deep-learning acceleration of non-local weighting and sparse inversion, and extension to differential TomoSAR (Shi et al., 2023). For automotive systems, suggested extensions include denser vertical arrays, non-vertical baselines, coded MIMO instead of TDM, generalized least-squares inversion for full 3D baseline geometry, advanced phase unwrapping, and multi-source fitting (Kabuli et al., 14 Jan 2025). For UAV swarms, the open problem is not only interferometric geometry but full sensing–communication co-design under real-time offloading constraints (Lahmeri et al., 2024, Lahmeri et al., 15 Jul 2025).

Taken together, these strands show that MB-InSAR is not a single algorithmic recipe but a family of inverse problems with shared phase geometry and highly variable statistical structure. Its unifying feature is the use of multiple baselines—temporal, spatial, or both—to turn interferometric phase from a pairwise observable into a redundant measurement system from which deformation histories, fused DEMs, or elevation spectra can be estimated with substantially greater stability and precision than single-baseline InSAR allows.