Abstract: Objectives： Synthetic Aperture Radar Interferometry (InSAR) relies heavily on the quality of the interferometric phase. However, in low-coherence regions, the phase is often severely corrupted by speckle noise and decorrelation, leading to broken fringe structures and widespread error propagation during subsequent phase unwrapping. Conventional Goldstein-type frequency-domain filters face significant limitations under these challenging conditions, including high-variance local spectral estimation, unstable noise-floor separation, and the generation of weak but problematic blocky artifacts caused by block-wise reconstruction strategies. The primary objective is to overcome these limitations by developing an advanced filtering architecture that balances heavy noise suppression with precise detail preservation. Methods： An enhanced filtering approach, termed the optimal-scale multi-taper and multi-offset lattice enhanced Goldstein filter (OS-MML-GF), is developed. To resolve the instability of local spectral estimation, multi-taper power spectral density (MPSD) estimation is introduced, which effectively smooths the random noise base. This is coupled with a robust noise-floor estimation and correction strategy to construct a highly stable frequency-domain filtering gain. Furthermore, to eliminate blocky artifacts, the method integrates a consistent overlap-add reconstruction technique with a multi-offset half-step lattice fusion mechanism. Processing is executed independently across multiple candidate spatial scales. Finally, an evaluation metric based on local phase gradient consistency (LGC) is formulated, enabling the pixel-wise dynamic selection of the multi-scale optimal filtering window to adaptively match the varying coherence levels across the interferogram. Results： Extensive evaluations were conducted on both simulated and real-world interferometric datasets. On three sets of simulated phase data featuring a coherence gradient ranging from 0.3 to 0.9, the proposed OS-MML-GF algorithm achieved a structural similarity index measure (SSIM) of up to 0.92, demonstrating superior fidelity. Furthermore, the number of residue points was drastically reduced to near zero, significantly outperforming conventional filters and lowering the subsequent phase unwrapping root-mean-square error to as low as 0.089 radians. In experiments utilizing real interferograms from the Myanmar earthquake and the Italian volcano regions, the algorithm effectively suppressed granular noise textures, enhanced both fringe continuity and boundary sharpness, and substantially increased the spatial pseudo-coherence distribution.. Conclusions： The developed OS-MML-GF algorithm successfully breaks through the physical limitations inherent in single-window-size filtering. It provides a highly robust balance between profound noise suppression in decorrelated areas and intricate detail preservation in highly coherent zones, ultimately supplying a highly reliable phase foundation for precision phase unwrapping and geophysical deformation inversion.