DRAPFNet: a dual-branch residual attention network for complex interferometric phase filtering in low-coherence airborne marine InSAR

Interferometric phase filtering is an important preprocessing step for coherence-degraded airborne marine interferometric synthetic aperture radar (InSAR), where decorrelation and rapid sea-surface evolution can induce strong wrapped-phase fluctuations and phase residues. This paper presents DRAPFNet, a dual-branch residual attention network that represents wrapped phase as a phase-only unit phasor with two real-valued channels. An encoder–decoder branch captures multi-scale spatial context, while a full-resolution mixed-dilation residual attention branch preserves locally rapid phase variations; inter-branch fusion and residual prediction combine both representations. Circular, complex-domain, and gradient-consistency losses are used to respect phase periodicity and local fringe structure. Evaluation includes coherence- and fringe-density-stratified simulations, a phase-unwrapping-oriented over-smoothing test, multi-seed statistical analysis, a three-moment 16.2-GHz airborne marine dataset, and an independent 9.6-GHz hybrid cross-/along-track airborne dataset. In the phase-unwrapping experiment, DRAPFNet achieves the lowest complex-gradient root-mean-square error, unwrapped-phase root-mean-square error, and large-error pixel ratio, while the Boxcar negative control shows that stronger smoothing can degrade the unwrapped phase. Across three training seeds, DRAPFNet attains the best mean wrapped-phase mean absolute error, peak signal-to-noise ratio, and structural similarity among the tested learning-based methods, with statistically significant paired differences. On measured data, DRAPFNet consistently reduces residue-related phase inconsistency; in the independent cross-/along-track experiment, residue density decreases by more than 99.8% in all four regions of interest. Because noise-free physical references are unavailable for the airborne observations, these real-data results are interpreted as no-reference evidence rather than direct validation of sea-surface phase or geophysical retrieval accuracy. Overall, DRAPFNet provides structure-aware phase filtering for the airborne marine conditions evaluated here.