Synthetic aperture radar, or SAR, produces complex images in which each pixel carries both amplitude and phase. Amplitude shows how strongly the ground reflected radar energy. Phase records where the radar wave is in its cycle at the moment of return. Most reconstruction and display methods ignore phase, treating it as a byproduct rather than an image-quality target. Yet for measuring how the ground moves between two satellite passes, phase difference is the entire observable. An arXiv preprint published on 2026-08-27 addresses this gap directly. [arXiv:2608.26605]
In interferometric SAR, or InSAR, two complex SAR acquisitions of the same area are combined. The phase difference between them is proportional to line-of-sight ground displacement. That is how InSAR monitors ground subsidence, landslides, and earthquake deformation. But sparse reconstruction methods have historically optimized magnitude fidelity, or single-epoch complex reconstruction accuracy. They do not optimize the differential phase that the interferogram actually uses. If two acquisitions are reconstructed independently, small phase errors in each can combine into large phase-difference errors. The challenge is to make two epochs share information where the scene is coherent but not force agreement where it has changed.
The Core Finding
The paper introduces the Differential-Phase-Oriented Joint Masked Reconstruction Network, DP-JMRNet. It uses deep unfolding to jointly reconstruct two SAR acquisitions from masked observations under an explicit differential-phase objective. Deep unfolding unrolls an iterative optimization into a sequence of learned network layers, preserving physical structure while learning parameters from data.
Instead of treating the two epochs separately, an exchange-equivariant interaction module makes the reconstruction order-invariant. A coherence-aware gate learns where to share information across epochs and where to close that sharing. On simulated bitemporal SAR data, DP-JMRNet reports the lowest differential-phase root mean squared error at 30%, 40%, and 50% sampling rates. The reduction is 47.5% to 51.3% against the best baseline, using one third of its parameters. The same trend appears on three Sentinel-1 scenes. A systematic acquisition design study shows that sharing the same aperture support across epochs is necessary for phase fidelity, while optimizing the sampling mask does not improve the differential phase.
DP-JMRNet attains the lowest differential-phase RMSE at 30%, 40%, and 50% sampling rate, while maintaining competitive amplitude and complex-image fidelity.
Think of it like two satellite photographs of a mountain slope. You don't care whether each image is perfectly exposed; you care whether a rock moved between the two moments. DP-JMRNet optimizes that movement measurement directly.
The State of the Field
Prior sparse reconstruction methods for SAR have focused on magnitude fidelity or single-epoch complex reconstruction. Deep unfolding networks in radar imaging typically unroll iterative algorithms and learn regularization parameters, but their loss functions rarely target interferometric phase difference. The result is a gap: a reconstruction that looks good as an image can still degrade the deformation estimate. This paper differs by making the differential phase itself the objective and by adding architectural constraints for epoch exchange and coherence-gated sharing.
Deep unfolding emerged as a way to combine the interpretability of iterative optimization with the flexibility of deep learning. In a typical unfolding network, each layer corresponds to one iteration of an algorithm such as iterative shrinkage-thresholding, or ISTA, and the network learns the shrinkage or regularization parameters from data. Here, the key change is that the objective is not a standard image reconstruction loss but a differential-phase loss. That change forces the network to preserve the interferometric quantity that downstream users actually need.
The broader Earth observation landscape is now shaped by freely available Sentinel-1 radar data, which since 2014 has created sustained demand for InSAR products that can monitor subsidence, landslides, and seismic deformation with fewer acquisitions. Sparse reconstruction matters because data gaps, downsampling, and onboard compression all limit the full complex data that reaches the ground segment. In this context, a network that preserves differential phase at low sampling rates addresses a bottleneck that is only growing as more SAR constellations produce more data.
From Lab to Reality
For scientists, the work unlocks a direct route from sparse SAR acquisitions to deformation products without first reconstructing magnitude-perfect images. It suggests that designing objectives around physical interferometric quantities, not just image similarity, can change what a network learns. The exchange-equivariant and coherence-aware modules are reusable architectural ideas for other multi-epoch remote sensing tasks.
For engineers, the code and data are available on GitHub, which may allow ground-segment processors to test phase-preserving reconstruction on Sentinel-1 and future missions such as NISAR. The method is not tied to a specific satellite and could be evaluated for upcoming SAR platforms where sampling is limited. For investors, this affects the InSAR analytics segment of Earth observation, where infrastructure monitoring, mining, and insurance sectors pay for millimeter-scale ground movement data. Operational adoption depends on validation across larger, real deformation events and integration into existing processing pipelines.
The practical deployment path may start with research processors rather than onboard satellites. Ground stations can already run deep networks on Sentinel-1 data, so the first commercial effect is likely to be in processing software rather than satellite hardware. If phase-preserving reconstruction reduces the amount of data required, it could also lower storage and transmission burdens for future SAR constellations.
What Still Needs to Happen
Several challenges remain. First, the paper validates the trend on three Sentinel-1 scenes, but a production InSAR processor must handle atmospheric phase screens, temporal decorrelation, and multi-year time series. Atmospheric water vapor introduces phase delays that can mimic ground motion. Temporal decorrelation occurs when surface scattering changes over time, especially in vegetated areas. These errors are not fully captured by simulated data or a small set of real scenes.
Second, the authors note that sharing the same aperture support across epochs is necessary for phase fidelity. That imposes an acquisition design constraint that satellite operators may need to adopt. The finding that optimizing the sampling mask does not improve differential phase suggests that future gains will come from network architecture and objective design, not from mask search. This is a useful negative result, but it also removes an easy lever. The DP-JMRNet authors, who have released code and data, are likely best positioned to extend the approach to real earthquake and subsidence cases. Other InSAR groups working with Sentinel-1 and NISAR will need to test whether the same coherence-aware sharing holds across varied terrain and deformation rates.
Another open issue is generalization. The paper reports a systematic study of acquisition design, but not a large benchmark across sensor modes, polarizations, or wavelengths. Future work will need to confirm whether the same differential-phase gains appear with L-band or X-band radar, where scattering physics differs.
Conclusion
In short, differential phase preservation changes what sparse bitemporal SAR reconstruction optimizes, cutting differential-phase RMSE by up to 51.3 percent against the best baseline. The result reframes phase from a diagnostic to a design objective.