Adapting InSAR Phase Linking for Seasonally Snow-Covered Terrain

Interferometric synthetic aperture radar (InSAR) time series analysis of natural terrain allows for characterization of long-term geophysical trends over extended areas and, in the case of distributed scatterers (DSs), is significantly enhanced by methods that exploit the full complex-valued scattering statistics. Phase-linking (PL) estimators impose a phase-closure constraint in order to estimate the temporal wrapped-phase history of a DS directly from its complex backscatter sample coherence matrix. Some PL methods, such as the SqueeSAR and maximum-likelihood-estimator of Interferometric phase (EMI) estimators, rely on knowledge of the coherence magnitude matrix. The true coherence magnitude is a priori unknown and must therefore be estimated from the data. Bias in these estimated coherence magnitudes reduces PL performance when the true coherence magnitude is low. Many areas of the Earth are seasonally snow-covered and, for natural terrain, this leads to severe cross-season decorrelation. This poses a significant challenge for PL estimators due to bias of the near-zero cross-season coherence magnitude estimates. We introduce a clustering approach to mitigate the PL estimator bias problem that exploits the fact that in natural terrain, many DSs decorrelate similarly. This allows for averaging over large numbers of same-behaving DS, which provides robust debiasing of the coherence magnitudes used during PL. We apply our method to a RADARSAT-2 spotlight-mode InSAR dataset over a site in the western Canadian Arctic and demonstrate significant reductions in a posteriori phase variance when compared to existing PL methods.

Automated summary

This paper introduces a clustering approach to mitigate the bias problem of phase-linking (PL) estimators in interferometric synthetic aperture radar (InSAR) time series analysis. The approach exploits the decorrelation similarity of distributed scatterers (DSs) in natural terrain for robust debiasing. Applying this method to a real dataset and comparing it with existing methods demonstrate significant improvements in performance.