Coseis‐Diff: A Data‐Independent Diffusion Transformer for Single Coseismic Interferogram Error Correction and Deformation Recovery

Abstract

Accurate retrieval of coseismic deformation fields from Interferometric Synthetic Aperture Radar is critical for constraining fault slip distributions and assessing seismic hazards. However, extracting pure deformation signals from single‐frame interferograms remains challenging due to the complex coupling of atmospheric turbulence, stratified delays, Digital Elevation Model residuals, orbital errors, and decorrelation noise, particularly in regions lacking external auxiliary data or extensive temporal archives. To address this, we propose Coseis‐Diff, an error‐correction model that requires no external auxiliary data at inference time, built upon a residual denoising diffusion framework. We employ a diffusion transformer backbone, which leverages global self‐attention mechanisms to overcome the limited receptive field of traditional Convolutional Neural Networks (CNNs). This architecture enables the unified correction of multi‐scale errors while preserving the topological continuity of deformation fringes. Evaluation on physics‐based synthetic data sets demonstrates that Coseis‐Diff achieves higher fidelity than CNN baselines. Furthermore, validation on real‐world earthquakes ( M _w 5.7–7.05) confirms that the model effectively separates spatially correlated artifacts and reliably recovers deformation signals even in severe low‐coherence zones. Comparisons against an independent Okada dislocation model and GNSS coseismic displacements further show that the recovered phase unwraps more reliably and agrees more closely with these independent observations than the raw interferogram. By replacing the reliance on external corrections or time‐series stacking with physics‐based learned priors, Coseis‐Diff provides a rapid, high‐precision solution for emergency response and source mechanism inversion in data‐scarce environments.